Merge remote-tracking branch 'origin/main' into litellm-providers/price-sync

litellm/__init__.py

@@ -538,7 +538,7 @@ context_window_fallbacks: Optional[List] = None
 content_policy_fallbacks: Optional[List] = None
 allowed_fails: int = 3
 allow_dynamic_callback_disabling: bool = True
-num_retries_per_request: Optional[int] = None  # for the request overall (incl. fallbacks + model retries)
+num_retries_per_request: Optional[int] = None  # cap on Router retries of one model group; resets per fallback hop
 ####### SECRET MANAGERS #####################
 secret_manager_client: Optional[Any] = (
     None  # list of instantiated key management clients - e.g. azure kv, infisical, etc.

litellm/_logging.py

@@ -1,5 +1,6 @@
 import ast
 import contextvars
+import functools
 import logging
 import os
 import sys
@@ -225,6 +226,35 @@ class AccessLogRedactionFilter(logging.Filter):
 _access_log_filter: Final = AccessLogRedactionFilter()
 
 
+@functools.lru_cache(maxsize=1)
+def _parse_disabled_access_log_paths(raw: str) -> frozenset[str]:
+    return frozenset(stripped for path in raw.split(",") if (stripped := path.strip()))
+
+
+def _disabled_access_log_paths() -> frozenset[str]:
+    """Read the variable per record so a value loaded later via proxy config
+    environment_variables or dotenv is honored."""
+    return _parse_disabled_access_log_paths(os.getenv("LITELLM_DISABLE_ACCESS_LOG_PATHS", ""))
+
+
+class AccessLogPathFilter(logging.Filter):
+    """Drops uvicorn.access records for request paths listed in LITELLM_DISABLE_ACCESS_LOG_PATHS.
+
+    uvicorn passes record.args as (client_addr, method, full_path, http_version, status_code).
+    """
+
+    def filter(self, record: logging.LogRecord) -> bool:
+        if not isinstance(record.args, tuple) or len(record.args) < 3:
+            return True
+        full_path: Final = record.args[2]
+        if not isinstance(full_path, str):
+            return True
+        return full_path.partition("?")[0] not in _disabled_access_log_paths()
+
+
+_access_log_path_filter: Final = AccessLogPathFilter()
+
+
 def _get_max_string_length_stdout_log() -> int:
     """Read the limit per record so a value loaded later via proxy config
     environment_variables is honored."""
@@ -663,6 +693,7 @@ def _redact_third_party_loggers() -> None:
     for name in _REDACTED_THIRD_PARTY_LOGGERS:
         logging.getLogger(name).addFilter(_secret_filter)
     for name in _REDACTED_ACCESS_LOGGERS:
+        logging.getLogger(name).addFilter(_access_log_path_filter)
         logging.getLogger(name).addFilter(_access_log_filter)
 
 

litellm/caching/caching.py

@@ -679,7 +679,14 @@ class Cache:
             cache_key, cached_data, kwargs = self._add_cache_logic(result=result, **kwargs)
             self.cache.set_cache(cache_key, cached_data, **kwargs)
         except Exception as e:
-            log_redis_failure(verbose_logger, logging.ERROR, "LiteLLM Cache: exception in add_cache", e)
+            self._log_add_cache_failure(e)
+
+    def _log_add_cache_failure(self, exc: Exception) -> None:
+        message: Final = "LiteLLM Cache: exception in add_cache"
+        if isinstance(self.cache, RedisCache):
+            log_redis_failure(verbose_logger, logging.ERROR, message, exc)
+            return
+        verbose_logger.error("%s: %s", message, exc)
 
     async def async_add_cache(self, result, dynamic_cache_object: BaseCache | None = None, **kwargs):
         """
@@ -698,7 +705,7 @@ class Cache:
                 else:
                     await self.cache.async_set_cache(cache_key, cached_data, **kwargs)
         except Exception as e:
-            log_redis_failure(verbose_logger, logging.ERROR, "LiteLLM Cache: exception in add_cache", e)
+            self._log_add_cache_failure(e)
 
     def _convert_to_cached_embedding(
         self,
@@ -877,7 +884,7 @@ class Cache:
             else:
                 await self.cache.async_set_cache_pipeline(cache_list=cache_list, **kwargs)
         except Exception as e:
-            log_redis_failure(verbose_logger, logging.ERROR, "LiteLLM Cache: exception in add_cache", e)
+            self._log_add_cache_failure(e)
 
     def should_use_cache(self, **kwargs):
         """

litellm/caching/redis_cache.py

@@ -15,6 +15,7 @@ import hashlib
 import inspect
 import json
 import logging
+import threading
 import time
 from collections.abc import Awaitable, Callable, Iterator, Sequence
 from contextvars import ContextVar
@@ -32,6 +33,7 @@ from litellm.constants import (
     REDIS_CIRCUIT_BREAKER_FAILURE_THRESHOLD,
     REDIS_CIRCUIT_BREAKER_RECOVERY_TIMEOUT,
     REDIS_CIRCUIT_BREAKER_TIMEOUT_MIN_DURATION,
+    REDIS_TIMEOUT_LOG_INTERVAL,
 )
 from litellm.litellm_core_utils.core_helpers import _get_parent_otel_span_from_kwargs
 from litellm.litellm_core_utils.coroutine_checker import coroutine_checker
@@ -340,7 +342,7 @@ def _explicit_causes(exc: BaseException) -> Iterator[BaseException]:
         current = current.__cause__
 
 
-def _is_redis_timeout_failure(exc: BaseException) -> bool:
+def is_redis_timeout_failure(exc: BaseException) -> bool:
     """True when ``exc`` or any exception it was explicitly raised ``from`` is a timeout.
 
     redis-py's blocking pool reports a pool wait timeout as ``ConnectionError`` chained from
@@ -414,7 +416,7 @@ def _record_swallowed_redis_failure(breaker: RedisCircuitBreaker, exc: BaseExcep
     """
     if not _is_redis_health_failure(exc):
         return
-    breaker.record_failure(is_timeout=_is_redis_timeout_failure(exc))
+    breaker.record_failure(is_timeout=is_redis_timeout_failure(exc))
     _swallowed_redis_failures.set(_swallowed_redis_failures.get() + 1)
 
 
@@ -422,13 +424,58 @@ class RedisCircuitBreakerOpenError(Exception):
     pass
 
 
+class _RedisTimeoutLogThrottle:
+    """Admits one Redis timeout log line per interval and counts the timeouts it suppressed in between."""
+
+    def __init__(self, interval: float, clock: Callable[[], float] = time.monotonic) -> None:
+        self.interval = interval
+        self._clock = clock
+        self._lock = threading.Lock()
+        self._last_logged_at: float | None = None
+        self._suppressed = 0
+
+    def admit(self) -> int | None:
+        """Return the number of timeouts suppressed since the last admitted line, or None to suppress this one."""
+        with self._lock:
+            now: Final = self._clock()
+            if self._last_logged_at is not None and now - self._last_logged_at < self.interval:
+                self._suppressed += 1
+                return None
+            suppressed: Final = self._suppressed
+            self._suppressed = 0
+            self._last_logged_at = now
+            return suppressed
+
+
+_redis_timeout_log_throttle: Final = _RedisTimeoutLogThrottle(REDIS_TIMEOUT_LOG_INTERVAL)
+
+
 def log_redis_failure(
     logger: logging.Logger, level: int, message: str, exc: BaseException, with_traceback: bool = False
 ) -> None:
     if isinstance(exc, RedisCircuitBreakerOpenError):
-        logger.debug("%s: %s", message, exc)
+        logger.debug("%s: %s", message, exc, stacklevel=2)
         return
-    logger.log(level, "%s: %s", message, exc, exc_info=exc if with_traceback else None)
+    exc_info: Final = exc if with_traceback else None
+    if not is_redis_timeout_failure(exc):
+        logger.log(level, "%s: %s", message, exc, exc_info=exc_info, stacklevel=2)
+        return
+    suppressed: Final = _redis_timeout_log_throttle.admit()
+    if suppressed is None:
+        logger.debug("%s: %s", message, exc, stacklevel=2)
+        return
+    if suppressed == 0:
+        logger.log(level, "%s: %s", message, exc, exc_info=exc_info, stacklevel=2)
+        return
+    logger.log(
+        level,
+        "%s: %s (%d more Redis timeouts since the previous Redis timeout line were logged at DEBUG)",
+        message,
+        exc,
+        suppressed,
+        exc_info=exc_info,
+        stacklevel=2,
+    )
 
 
 @dataclass(frozen=True, slots=True)
@@ -475,7 +522,7 @@ async def _run_under_circuit_breaker(
         result: Final = await call()
     except Exception as e:
         if _is_redis_health_failure(e):
-            breaker.record_failure(is_timeout=_is_redis_timeout_failure(e))
+            breaker.record_failure(is_timeout=is_redis_timeout_failure(e))
         raise
     _exit_circuit_breaker(breaker, admission)
     return result
@@ -492,7 +539,7 @@ def _run_under_circuit_breaker_sync(
         result: Final = call()
     except Exception as e:
         if _is_redis_health_failure(e):
-            breaker.record_failure(is_timeout=_is_redis_timeout_failure(e))
+            breaker.record_failure(is_timeout=is_redis_timeout_failure(e))
         raise
     _exit_circuit_breaker(breaker, admission)
     return result
@@ -801,10 +848,8 @@ class RedisCache(BaseCache):
             ## LOGGING ##
             end_time = time.time()
             _duration = end_time - start_time
-            verbose_logger.error(
-                "LiteLLM Redis Caching: increment_cache() - Got exception from REDIS %s, Writing value=%s",
-                str(e),
-                value,
+            log_redis_failure(
+                verbose_logger, logging.ERROR, "LiteLLM Redis Caching: increment_cache() - Got exception from REDIS", e
             )
             raise e
 
@@ -1010,11 +1055,8 @@ class RedisCache(BaseCache):
                     call_type=f"async_set_cache <- {_get_call_stack_info()}",
                 )
             )
-            verbose_logger.error(
-                "LiteLLM Redis Caching: async set() - Got exception from REDIS %s, key=%r, value=%r",
-                str(e),
-                key,
-                value,
+            log_redis_failure(
+                verbose_logger, logging.ERROR, "LiteLLM Redis Caching: async set() - Got exception from REDIS", e
             )
             raise e
 
@@ -1062,10 +1104,8 @@ class RedisCache(BaseCache):
                     event_metadata={"key": key},
                 )
             )
-            verbose_logger.error(
-                "LiteLLM Redis Caching: async set() - Got exception from REDIS %s, Writing value=%s",
-                str(e),
-                value,
+            log_redis_failure(
+                verbose_logger, logging.ERROR, "LiteLLM Redis Caching: async set() - Got exception from REDIS", e
             )
             _record_swallowed_redis_failure(self._circuit_breaker, e)
 
@@ -1112,7 +1152,6 @@ class RedisCache(BaseCache):
         start_time: Final = time.time()
 
         print_verbose(f"Set Async Redis Cache: key list: {cache_list}\nttl={ttl}, redis_version={self.redis_version}")
-        cache_value: Final = None
         try:
             async with _redis_client.pipeline(transaction=False) as pipe:
                 results: Final = await self._pipeline_helper(pipe, cache_list, ttl)
@@ -1149,10 +1188,11 @@ class RedisCache(BaseCache):
                 )
             )
 
-            verbose_logger.error(
-                "LiteLLM Redis Caching: async set_cache_pipeline() - Got exception from REDIS %s, Writing value=%s",
-                str(e),
-                cache_value,
+            log_redis_failure(
+                verbose_logger,
+                logging.ERROR,
+                "LiteLLM Redis Caching: async set_cache_pipeline() - Got exception from REDIS",
+                e,
             )
             _record_swallowed_redis_failure(self._circuit_breaker, e)
 
@@ -1191,8 +1231,11 @@ class RedisCache(BaseCache):
                     end_time=time.time(),
                 )
             )
-            verbose_logger.error(
-                "LiteLLM Redis Caching: async_set_cache_pipeline_with_ttls() - Got exception from REDIS %s", str(e)
+            log_redis_failure(
+                verbose_logger,
+                logging.ERROR,
+                "LiteLLM Redis Caching: async_set_cache_pipeline_with_ttls() - Got exception from REDIS",
+                e,
             )
             _record_swallowed_redis_failure(self._circuit_breaker, e)
 
@@ -1235,10 +1278,8 @@ class RedisCache(BaseCache):
                 )
             )
             # NON blocking - notify users Redis is throwing an exception
-            verbose_logger.error(
-                "LiteLLM Redis Caching: async set() - Got exception from REDIS %s, Writing value=%s",
-                str(e),
-                value,
+            log_redis_failure(
+                verbose_logger, logging.ERROR, "LiteLLM Redis Caching: async set() - Got exception from REDIS", e
             )
             raise e
 
@@ -1274,10 +1315,11 @@ class RedisCache(BaseCache):
                 )
             )
             # NON blocking - notify users Redis is throwing an exception
-            verbose_logger.error(
-                "LiteLLM Redis Caching: async set_cache_sadd() - Got exception from REDIS %s, Writing value=%s",
-                str(e),
-                value,
+            log_redis_failure(
+                verbose_logger,
+                logging.ERROR,
+                "LiteLLM Redis Caching: async set_cache_sadd() - Got exception from REDIS",
+                e,
             )
             _record_swallowed_redis_failure(self._circuit_breaker, e)
 
@@ -1359,10 +1401,11 @@ class RedisCache(BaseCache):
                     parent_otel_span=parent_otel_span,
                 )
             )
-            verbose_logger.error(
-                "LiteLLM Redis Caching: async async_increment() - Got exception from REDIS %s, Writing value=%s",
-                str(e),
-                value,
+            log_redis_failure(
+                verbose_logger,
+                logging.ERROR,
+                "LiteLLM Redis Caching: async async_increment() - Got exception from REDIS",
+                e,
             )
             raise e
 
@@ -1448,7 +1491,9 @@ class RedisCache(BaseCache):
             print_verbose(f"Got Redis Cache: key: {key}, cached_response {cached_response}")
             return self._get_cache_logic(cached_response=cached_response)
         except Exception as e:
-            verbose_logger.error("litellm.caching.caching: get() - Got exception from REDIS: %s", e)
+            log_redis_failure(
+                verbose_logger, logging.ERROR, "litellm.caching.caching: get() - Got exception from REDIS", e
+            )
             _record_swallowed_redis_failure(self._circuit_breaker, e)
 
     def _run_redis_mget_operation(self, keys: list[str]) -> Sequence[bytes | str | None]:
@@ -1526,7 +1571,7 @@ class RedisCache(BaseCache):
                 end_time=failed_at,
                 parent_otel_span=parent_otel_span,
             )
-            verbose_logger.error("Error occurred in batch get cache - %s", e)
+            log_redis_failure(verbose_logger, logging.ERROR, "Error occurred in batch get cache", e)
             _record_swallowed_redis_failure(self._circuit_breaker, e)
             return key_value_dict
 
@@ -1645,7 +1690,7 @@ class RedisCache(BaseCache):
                     parent_otel_span=parent_otel_span,
                 )
             )
-            verbose_logger.error("Error occurred in async batch get cache - %s", e)
+            log_redis_failure(verbose_logger, logging.ERROR, "Error occurred in async batch get cache", e)
             _record_swallowed_redis_failure(self._circuit_breaker, e)
             return key_value_dict
 
@@ -1870,9 +1915,11 @@ class RedisCache(BaseCache):
                     parent_otel_span=_get_parent_otel_span_from_kwargs(kwargs),
                 )
             )
-            verbose_logger.error(
-                "LiteLLM Redis Caching: async increment_pipeline() - Got exception from REDIS %s",
-                str(e),
+            log_redis_failure(
+                verbose_logger,
+                logging.ERROR,
+                "LiteLLM Redis Caching: async increment_pipeline() - Got exception from REDIS",
+                e,
             )
             raise e
 
@@ -1949,7 +1996,7 @@ class RedisCache(BaseCache):
                     call_type=f"async_rpush <- {_get_call_stack_info()}",
                 )
             )
-            verbose_logger.error("LiteLLM Redis Cache RPUSH: - Got exception from REDIS : %s", e)
+            log_redis_failure(verbose_logger, logging.ERROR, "LiteLLM Redis Cache RPUSH: - Got exception from REDIS", e)
             raise e
 
     async def _pipeline_rpush_helper(
@@ -2017,9 +2064,11 @@ class RedisCache(BaseCache):
                     call_type=f"async_rpush_pipeline <- {_get_call_stack_info()}",
                 )
             )
-            verbose_logger.error(
-                "LiteLLM Redis Caching: async_rpush_pipeline() - Got exception from REDIS %s",
-                str(e),
+            log_redis_failure(
+                verbose_logger,
+                logging.ERROR,
+                "LiteLLM Redis Caching: async_rpush_pipeline() - Got exception from REDIS",
+                e,
             )
             raise e
 
@@ -2095,7 +2144,7 @@ class RedisCache(BaseCache):
                     call_type=f"async_lpop <- {_get_call_stack_info()}",
                 )
             )
-            verbose_logger.error("LiteLLM Redis Cache LPOP: - Got exception from REDIS : %s", e)
+            log_redis_failure(verbose_logger, logging.ERROR, "LiteLLM Redis Cache LPOP: - Got exception from REDIS", e)
             raise e
 
     async def _pipeline_lpop_helper(
@@ -2206,8 +2255,10 @@ class RedisCache(BaseCache):
                     call_type=f"async_lpop_pipeline <- {_get_call_stack_info()}",
                 )
             )
-            verbose_logger.error(
-                "LiteLLM Redis Caching: async_lpop_pipeline() - Got exception from REDIS %s",
-                str(e),
+            log_redis_failure(
+                verbose_logger,
+                logging.ERROR,
+                "LiteLLM Redis Caching: async_lpop_pipeline() - Got exception from REDIS",
+                e,
             )
             raise e

litellm/constants.py

@@ -227,6 +227,9 @@ PRE_CALL_EXECUTED_GUARDRAILS_KEY: Final = "_pre_call_executed_guardrails"
 # Attribute stamped on log_guardrail_information wrappers so __init_subclass__ does not wrap them again
 LOGS_GUARDRAIL_INFORMATION_MARKER: Final = "_litellm_logs_guardrail_information"
 
+# llm_provider stamped on proxy-side rate limit errors when the model resolves to no deployment
+PROXY_LLM_PROVIDER_FALLBACK: Final = "litellm_proxy"
+
 # Generic fallback for unknown models
 DEFAULT_REASONING_EFFORT_MINIMAL_THINKING_BUDGET: Final = int(
     os.getenv("DEFAULT_REASONING_EFFORT_MINIMAL_THINKING_BUDGET", 128)
@@ -311,6 +314,10 @@ REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS: Final = float(
 # RFC 6455 caps the close frame payload at 125 bytes, 2 of which carry the status code
 WEBSOCKET_CLOSE_REASON_MAX_BYTES: Final = 123
 
+BEDROCK_REALTIME_PENDING_SESSION_UPDATE_SCOPE_KEY: Final = "litellm.bedrock_realtime.pending_session_update"
+BEDROCK_REALTIME_SESSION_COMMITTED_SCOPE_KEY: Final = "litellm.bedrock_realtime.session_committed"
+BEDROCK_REALTIME_COMMITTED_FAILURE_SCOPE_KEY: Final = "litellm.bedrock_realtime.committed_failure"
+
 # SSL/TLS cipher configuration for faster handshakes
 # Strategy: Strongly prefer fast modern ciphers, but allow fallback to commonly supported ones
 # This balances performance with broad compatibility
@@ -461,6 +468,7 @@ REDIS_CIRCUIT_BREAKER_ENABLED: Final = os.getenv("REDIS_CIRCUIT_BREAKER_ENABLED"
 # minimum seconds a timeout-only failure streak must span before it can open the breaker,
 # so one event-loop stall timing out many queued calls at once does not trip it
 REDIS_CIRCUIT_BREAKER_TIMEOUT_MIN_DURATION: Final = float(os.getenv("REDIS_CIRCUIT_BREAKER_TIMEOUT_MIN_DURATION", 5.0))
+REDIS_TIMEOUT_LOG_INTERVAL: Final = float(os.getenv("REDIS_TIMEOUT_LOG_INTERVAL", "5.0"))
 # Seconds of idle before a Redis cluster connection is validated with a PING and
 # reconnected if dead, so a connection silently dropped by a cluster restart
 # (e.g. ElastiCache Serverless maintenance) is not reused while broken

litellm/cost_calculator.py

@@ -97,6 +97,7 @@ from litellm.llms.vertex_ai.cost_calculator import cost_router as google_cost_ro
 from litellm.llms.xai.cost_calculator import cost_per_token as xai_cost_per_token
 from litellm.responses.utils import ResponseAPILoggingUtils
 from litellm.types.agents import LiteLLMSendMessageResponse
+from litellm.types.llms.base import CachedTokensDetails
 from litellm.types.llms.openai import (
     HttpxBinaryResponseContent,
     ImageGenerationRequestQuality,
@@ -2381,6 +2382,46 @@ def _summable_prompt_token_fields(prompt_tokens_details: BaseModel) -> list[str]
     return [attr for attr in field_names if attr != "cache_creation_tokens"]
 
 
+def _combine_cached_tokens_details(
+    current: CachedTokensDetails | None, new: CachedTokensDetails
+) -> CachedTokensDetails:
+    def _sum_optional(current_value: int | None, new_value: int | None) -> int | None:
+        if current_value is None and new_value is None:
+            return None
+        return (current_value or 0) + (new_value or 0)
+
+    return CachedTokensDetails(
+        text_tokens=_sum_optional(current.text_tokens if current is not None else None, new.text_tokens),
+        audio_tokens=_sum_optional(current.audio_tokens if current is not None else None, new.audio_tokens),
+        image_tokens=_sum_optional(current.image_tokens if current is not None else None, new.image_tokens),
+    )
+
+
+def _combine_prompt_tokens_details(
+    current: PromptTokensDetailsWrapper | None, new: PromptTokensDetailsWrapper
+) -> PromptTokensDetailsWrapper:
+    base: Final = current if current is not None else PromptTokensDetailsWrapper()
+    base_values: Final = MappingProxyType(
+        {attr: getattr(base, attr) for attr in type(base).model_fields if hasattr(base, attr)}
+    )
+    summed: Final = MappingProxyType(
+        {
+            attr: (getattr(base, attr, 0) or 0) + (getattr(new, attr) or 0)
+            for attr in _summable_prompt_token_fields(new)
+            if hasattr(new, attr) and isinstance(getattr(new, attr) or 0, (int, float))
+        }
+    )
+    new_cached_tokens_details: Final = getattr(new, "cached_tokens_details", None)
+    cached_tokens_details: Final = (
+        _combine_cached_tokens_details(getattr(base, "cached_tokens_details", None), new_cached_tokens_details)
+        if isinstance(new_cached_tokens_details, CachedTokensDetails)
+        else getattr(base, "cached_tokens_details", None)
+    )
+    return PromptTokensDetailsWrapper(
+        **MappingProxyType({**base_values, **summed, "cached_tokens_details": cached_tokens_details})
+    )
+
+
 class BaseTokenUsageProcessor:
     @staticmethod
     def combine_usage_objects(usage_objects: list[Usage]) -> Usage:
@@ -2389,7 +2430,6 @@ class BaseTokenUsageProcessor:
         """
         from litellm.types.utils import (
             CompletionTokensDetailsWrapper,
-            PromptTokensDetailsWrapper,
             Usage,
         )
 
@@ -2408,27 +2448,10 @@ class BaseTokenUsageProcessor:
                         and isinstance(current_val, (int, float))
                     ):
                         setattr(combined, attr, current_val + new_val)
-            # Handle nested prompt_tokens_details
             if hasattr(usage, "prompt_tokens_details") and usage.prompt_tokens_details:
-                if not hasattr(combined, "prompt_tokens_details") or not combined.prompt_tokens_details:
-                    combined.prompt_tokens_details = PromptTokensDetailsWrapper()
-
-                # Check what keys exist in the model's prompt_tokens_details
-                # Access model_fields on the class, not the instance, to avoid Pydantic 2.11+ deprecation warnings
-                for attr in _summable_prompt_token_fields(usage.prompt_tokens_details):
-                    if (
-                        hasattr(usage.prompt_tokens_details, attr)
-                        and not attr.startswith("_")
-                        and not callable(_attribute_value(usage.prompt_tokens_details, attr))
-                    ):
-                        current_val = getattr(combined.prompt_tokens_details, attr, 0) or 0
-                        new_val = getattr(usage.prompt_tokens_details, attr, 0) or 0
-                        if new_val is not None and isinstance(new_val, (int, float)):
-                            setattr(
-                                combined.prompt_tokens_details,
-                                attr,
-                                current_val + new_val,
-                            )
+                combined.prompt_tokens_details = _combine_prompt_tokens_details(
+                    getattr(combined, "prompt_tokens_details", None), usage.prompt_tokens_details
+                )
 
             # Handle nested completion_tokens_details
             if hasattr(usage, "completion_tokens_details") and usage.completion_tokens_details:

litellm/integrations/otel/mappers/openinference.py

@@ -12,6 +12,7 @@ from typing import Final
 
 from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue, SpanData
 from litellm.integrations.otel.mappers.utils import (
+    MAX_MESSAGE_ATTRS_PER_SPAN,
     MAX_TOOL_DEFINITION_ATTRS_PER_SPAN,
     collect,
     drop_none,
@@ -26,6 +27,8 @@ from litellm.integrations.otel.model.payloads import (
     ToolDefinition,
 )
 
+_MAX_INDEXED_MESSAGES: Final = MAX_MESSAGE_ATTRS_PER_SPAN // 2
+
 
 class OpenInferenceMapper:
     """Emits OpenInference attributes for LLM_CALL spans.
@@ -84,27 +87,44 @@ class OpenInferenceMapper:
                 return {}
 
     def _llm_call(self, data: LLMCallSpanData) -> AttributeMap:
+        outputs: Final = output_messages(data)
+        indexed_in, indexed_out = self._indexed_split(len(data.messages_in), len(outputs))
         return {
             **collect(self._LLM_CALL_ATTRS, data),
             **collect(self._BLOB_ATTRS, data),
-            **self._messages("llm.input_messages", "input.value", data.messages_in),
-            **self._messages("llm.output_messages", "output.value", output_messages(data)),
+            **self._messages(
+                "llm.input_messages",
+                "input.value",
+                data.messages_in,
+                self._prompt_positions(len(data.messages_in), indexed_in),
+            ),
+            **self._messages("llm.output_messages", "output.value", outputs, range(indexed_out)),
             **self._tools(data),
         }
 
     @staticmethod
-    def _messages(prefix: str, value_key: str, messages: Sequence[object]) -> AttributeMap:
-        """Per-message ``{prefix}.{idx}.message.*`` keys + the ``value_key`` blob."""
+    def _indexed_split(inputs: int, outputs: int) -> tuple[int, int]:
+        """Prompt and response share one allowance; the response is reserved at least half of it."""
+        indexed_out: Final = min(outputs, max(_MAX_INDEXED_MESSAGES // 2, _MAX_INDEXED_MESSAGES - inputs))
+        return _MAX_INDEXED_MESSAGES - indexed_out, indexed_out
+
+    @staticmethod
+    def _prompt_positions(total: int, indexed: int) -> tuple[int, ...]:
+        """Prompt messages that get per-index attributes: message 0 and the most recent turns."""
+        if total <= indexed:
+            return tuple(range(total))
+        return (0, *range(total - indexed + 1, total))
+
+    @staticmethod
+    def _messages(prefix: str, value_key: str, messages: Sequence[object], positions: Sequence[int]) -> AttributeMap:
+        """``{prefix}.{idx}.message.*`` keys for the messages at ``positions`` + the ``value_key`` blob of all."""
         parsed: Final = [(m.get("role") if isinstance(m, dict) else None, message_content(m)) for m in messages]
         attrs: Final = drop_none(
             {
                 key: value
-                for idx, (role, content) in enumerate(parsed)
+                for idx, (role, content) in ((idx, parsed[idx]) for idx in positions)
                 for key, value in (
-                    (
-                        f"{prefix}.{idx}.message.role",
-                        role if isinstance(role, str) else None,
-                    ),
+                    (f"{prefix}.{idx}.message.role", role if isinstance(role, str) else None),
                     (f"{prefix}.{idx}.message.content", content),
                 )
             }

litellm/integrations/otel/mappers/utils.py

@@ -32,6 +32,14 @@ core telemetry no matter how many vocabularies are configured.
 """
 
 
+MAX_MESSAGE_ATTRS_PER_SPAN: Final = DEFAULT_SPAN_ATTRIBUTE_LIMIT // 8
+"""Span-wide ceiling on per-index chat message attributes, prompt and response together.
+
+An eighth is the largest share that still fits beside the tool ceiling and the core
+of every vocabulary at once. The complete conversation still rides the JSON blobs.
+"""
+
+
 def tool_attr_budget(vocabularies: int) -> int:
     """Split the span-wide tool-definition ceiling across active vocabularies."""
     return MAX_TOOL_DEFINITION_ATTRS_PER_SPAN // max(vocabularies, 1)

litellm/integrations/prometheus.py

@@ -16,6 +16,7 @@ from pydantic import BaseModel
 
 import litellm
 from litellm._logging import print_verbose, verbose_logger
+from litellm.constants import PROXY_LLM_PROVIDER_FALLBACK
 from litellm.exceptions import (
     validate_rate_limit_category,
     validate_rate_limit_type,
@@ -2581,6 +2582,15 @@ class PrometheusLogger(CustomLogger):
             )
             return None
 
+    @staticmethod
+    def _extract_api_provider_from_exception(exception: Exception) -> str | None:
+        if not isinstance(exception, litellm.exceptions.RateLimitError):
+            return None
+        llm_provider: Final = exception.llm_provider
+        if not llm_provider or llm_provider == PROXY_LLM_PROVIDER_FALLBACK:
+            return None
+        return llm_provider
+
     async def async_post_call_failure_hook(
         self,
         request_data: dict,
@@ -2616,7 +2626,9 @@ class PrometheusLogger(CustomLogger):
             _metadata: Final = request_data.get("metadata", {}) or {}
             model_id: Final = _metadata.get("model_info", {}).get("id") or request_data.get("model_info", {}).get("id")
             rate_limit_category, rate_limit_type = self._extract_rate_limit_labels(original_exception)
-            api_provider: Final = self._extract_api_provider_from_request_data(request_data)
+            api_provider: Final = self._extract_api_provider_from_request_data(
+                request_data
+            ) or self._extract_api_provider_from_exception(original_exception)
             enum_values: Final = UserAPIKeyLabelValues(
                 end_user=user_api_key_dict.end_user_id,
                 user=user_api_key_dict.user_id,

litellm/litellm_core_utils/core_helpers.py

@@ -303,6 +303,16 @@ def get_metadata_variable_name_from_kwargs(
     return "litellm_metadata" if "litellm_metadata" in kwargs else "metadata"
 
 
+def max_retries_per_request_hit(kwargs: Mapping[str, object], num_retries_per_request: int | None) -> bool:
+    if num_retries_per_request is None:
+        return False
+    metadata: Final = kwargs.get(get_metadata_variable_name_from_kwargs(kwargs))
+    if not isinstance(metadata, Mapping):
+        return False
+    attempted_retries: Final = metadata.get("attempted_retries")
+    return type(attempted_retries) is int and 0 < attempted_retries and num_retries_per_request <= attempted_retries
+
+
 def get_or_create_metadata_bucket(
     request_data: dict,
 ) -> tuple[Literal["metadata", "litellm_metadata"], dict]:

litellm/litellm_core_utils/fallback_generalizations.py

@@ -34,6 +34,11 @@ rules never mix the two and never use ``extends``. A rule whose
 Rules are only consulted after exact and case-insensitive lookups miss, so an
 exact cost-map entry always takes precedence over any rule.
 
+Rules flagged with ``fill_missing_for_providers: [..]`` also fill only keys
+missing from an exact cost-map entry when the entry's ``litellm_provider`` is
+listed, while values already present on the entry win on conflict. Only flagged
+capability rules participate in this fill; routing rules never do.
+
 Patterns are matched case-insensitively with ``re.search`` and are not implicitly
 anchored: a rule must include ``^`` and ``$`` to bind to the whole model name,
 otherwise it matches as a substring. Keeping anchoring in the regex makes the rule
@@ -46,17 +51,19 @@ Rules are compiled and classified once, at install time. The match functions are
 O(number of rules); callers must only invoke them on a cache miss.
 """
 
+import logging
 import re
+from collections.abc import Mapping
 from dataclasses import dataclass
 from typing import Final
 
-from litellm._logging import verbose_logger
-
+verbose_logger: Final = logging.getLogger("LiteLLM")
 NAME_FIELD: Final = "name"
 PATTERN_FIELD: Final = "pattern"
 MODEL_INFO_FIELD: Final = "model_info"
 PROVIDER_KEY: Final = "litellm_provider"
 LEGACY_EXTENDS_FIELD: Final = "extends"
+FILL_MISSING_FOR_PROVIDERS_FIELD: Final = "fill_missing_for_providers"
 
 
 def _resolve_legacy_extends(rules: list) -> list:
@@ -98,11 +105,28 @@ class _RoutingRule:
 class _CapabilityRule:
     pattern: re.Pattern
     model_info: dict
+    fill_missing_for_providers: frozenset[str]
 
 
 _CompiledRule = _RoutingRule | _CapabilityRule
 
 
+def _parse_fill_missing_for_providers(rule: Mapping[str, object], pattern_label: object) -> frozenset[str] | None:
+    if FILL_MISSING_FOR_PROVIDERS_FIELD not in rule:
+        return frozenset()
+    raw_fill_missing_for_providers: Final = rule.get(FILL_MISSING_FOR_PROVIDERS_FIELD)
+    if not isinstance(raw_fill_missing_for_providers, (list, tuple)) or not all(
+        isinstance(provider, str) for provider in raw_fill_missing_for_providers
+    ):
+        verbose_logger.warning(
+            "LiteLLM: skipping malformed fallback generalization rule %s ('%s' must be a list of provider strings).",
+            rule.get(NAME_FIELD, pattern_label),
+            FILL_MISSING_FOR_PROVIDERS_FIELD,
+        )
+        return None
+    return frozenset(raw_fill_missing_for_providers)
+
+
 def _compile_rule(rule: object) -> tuple[_CompiledRule, ...]:
     if not isinstance(rule, dict):
         return ()
@@ -125,8 +149,17 @@ def _compile_rule(rule: object) -> tuple[_CompiledRule, ...]:
             e,
         )
         return ()
+    fill_missing_for_providers: Final = _parse_fill_missing_for_providers(rule, pattern)
+    if fill_missing_for_providers is None:
+        return ()
     if PROVIDER_KEY not in model_info:
-        return (_CapabilityRule(pattern=compiled, model_info=model_info),)
+        return (
+            _CapabilityRule(
+                pattern=compiled,
+                model_info=model_info,
+                fill_missing_for_providers=fill_missing_for_providers,
+            ),
+        )
     provider: Final = model_info[PROVIDER_KEY]
     if not isinstance(provider, str):
         verbose_logger.warning(
@@ -140,7 +173,11 @@ def _compile_rule(rule: object) -> tuple[_CompiledRule, ...]:
         return (_RoutingRule(pattern=compiled, provider=provider),)
     return (
         _RoutingRule(pattern=compiled, provider=provider),
-        _CapabilityRule(pattern=compiled, model_info=model_info),
+        _CapabilityRule(
+            pattern=compiled,
+            model_info=model_info,
+            fill_missing_for_providers=fill_missing_for_providers,
+        ),
     )
 
 
@@ -151,6 +188,7 @@ class _FallbackGeneralizations:
         self.rules: list = []
         self.routing_rules: tuple = ()
         self.capability_rules: tuple = ()
+        self.fill_missing_rules: tuple[_CapabilityRule, ...] = ()
 
     def set_rules(self, rules: list | None) -> None:
         installed: Final = rules if isinstance(rules, list) else []
@@ -158,6 +196,7 @@ class _FallbackGeneralizations:
         self.rules = installed
         self.routing_rules = tuple(rule for rule in compiled if isinstance(rule, _RoutingRule))
         self.capability_rules = tuple(rule for rule in compiled if isinstance(rule, _CapabilityRule))
+        self.fill_missing_rules = tuple(rule for rule in self.capability_rules if rule.fill_missing_for_providers)
 
     def match_routing(self, model: str) -> str | None:
         if not model:
@@ -175,6 +214,21 @@ class _FallbackGeneralizations:
             return None
         return {key: value for model_info in matched for key, value in model_info.items()}
 
+    def match_fill_missing(self, model: str, provider: str) -> Mapping[str, object] | None:
+        if not model or not provider:
+            return None
+        matched = tuple(
+            rule.model_info
+            for rule in self.fill_missing_rules
+            if provider in rule.fill_missing_for_providers and rule.pattern.search(model) is not None
+        )
+        if not matched:
+            return None
+        fill_missing: Final[Mapping[str, object]] = {
+            key: value for model_info in matched for key, value in model_info.items() if key != PROVIDER_KEY
+        }
+        return fill_missing or None
+
 
 _registry: Final = _FallbackGeneralizations()
 
@@ -210,3 +264,14 @@ def match_capability_generalizations(model: str) -> dict | None:
     capability rule matches. O(number of rules); only call once exact lookups have missed.
     """
     return _registry.match_capabilities(model)
+
+
+def match_fill_missing_generalizations(model: str, provider: str) -> Mapping[str, object] | None:
+    """Return flagged capability rules matching ``model`` for ``provider``.
+
+    Later rules override earlier ones on key conflicts. Only rules listing
+    ``provider`` in ``fill_missing_for_providers`` contribute. Returns ``None``
+    when no flagged rule matches. O(number of rules); only call once exact
+    lookups have matched.
+    """
+    return _registry.match_fill_missing(model, provider)

litellm/litellm_core_utils/litellm_logging.py

@@ -556,7 +556,7 @@ class Logging(LiteLLMLoggingBaseClass):
         # ids leaking into a different, later request on the same thread. Sync
         # support is deferred to a follow-up PR with its own safe-restore
         # mechanism; async calls (the proxy's only call path) are unaffected.
-        if supports_correlation_logging:
+        if supports_correlation_logging and litellm.request_correlation_in_logs:
             set_trace_id(self.litellm_trace_id)
             set_session_id(self.litellm_session_id)
         # set_trace_id()/set_session_id() sanitize (strip control chars, bound
@@ -2442,7 +2442,7 @@ class Logging(LiteLLMLoggingBaseClass):
         call) would leave the outer request's subsequent log lines stamped with
         the nested call's trace_id/session_id instead of its own.
 
-        Uses a plain set() of the captured pre-call value rather than
+        Uses a plain contextvar set() of the captured pre-call value rather than
         contextvars.Token-based reset(), since this can end up called from a
         different asyncio Task/context than __init__ ran in (e.g. the request
         task's own wrapper() finally block, plus async_success_handler
@@ -2453,8 +2453,8 @@ class Logging(LiteLLMLoggingBaseClass):
         that Task's view of the contextvars, so calling it multiple times
         (once per Task involved in this attempt) is required, not just safe.
         """
-        set_trace_id(self._pre_call_trace_id)
-        set_session_id(self._pre_call_session_id)
+        trace_id_var.set(self._pre_call_trace_id)
+        session_id_var.set(self._pre_call_session_id)
 
     def _restore_correlation_context_if_unclaimed(self) -> None:
         """Guarded variant for __del__-triggered cleanup only.

litellm/litellm_core_utils/llm_cost_calc/utils.py

@@ -9,6 +9,8 @@ from types import MappingProxyType
 from typing import Any, Final, Literal, TypedDict, cast
 from zoneinfo import ZoneInfo, ZoneInfoNotFoundError
 
+from typing_extensions import ReadOnly
+
 import litellm
 from litellm._internal_context import current_billing_time
 from litellm._logging import verbose_logger
@@ -772,6 +774,7 @@ def calculate_cache_writing_cost(
 
 class PromptTokensDetailsResult(TypedDict):
     cache_hit_tokens: int
+    cache_hit_audio_tokens: ReadOnly[int]
     cache_creation_tokens: int
     cache_creation_token_details: CacheCreationTokenDetails | None
     text_tokens: int
@@ -802,12 +805,34 @@ def parse_prompt_tokens_details(usage: Usage) -> PromptTokensDetailsResult:
         )
         or None
     )
-    text_tokens: Final = (
-        cast(int | None, getattr(usage.prompt_tokens_details, "text_tokens", None))
-        or 0  # default to prompt tokens, if this field is not set
+    cached_tokens_details: Final = getattr(usage.prompt_tokens_details, "cached_tokens_details", None)
+    cached_audio_tokens: Final = min(
+        _get_token_detail_value(cached_tokens_details, "audio_tokens") or 0, cache_hit_tokens
+    )
+    cached_text_tokens: Final = min(
+        _get_token_detail_value(cached_tokens_details, "text_tokens") or 0,
+        cache_hit_tokens - cached_audio_tokens,
+    )
+    cached_image_tokens: Final = min(
+        _get_token_detail_value(cached_tokens_details, "image_tokens") or 0,
+        cache_hit_tokens - cached_audio_tokens - cached_text_tokens,
+    )
+    text_tokens: Final = max(
+        (
+            cast(int | None, getattr(usage.prompt_tokens_details, "text_tokens", None))
+            or 0  # default to prompt tokens, if this field is not set
+        )
+        - cached_text_tokens,
+        0,
+    )
+    audio_tokens: Final = max(
+        (cast(int | None, getattr(usage.prompt_tokens_details, "audio_tokens", 0)) or 0) - cached_audio_tokens,
+        0,
+    )
+    image_tokens: Final = max(
+        (cast(int | None, getattr(usage.prompt_tokens_details, "image_tokens", 0)) or 0) - cached_image_tokens,
+        0,
     )
-    audio_tokens: Final = cast(int | None, getattr(usage.prompt_tokens_details, "audio_tokens", 0)) or 0
-    image_tokens: Final = cast(int | None, getattr(usage.prompt_tokens_details, "image_tokens", 0)) or 0
     video_tokens: Final = _coerce_token_count(getattr(usage.prompt_tokens_details, "video_tokens", 0))
     character_count: Final = (
         cast(
@@ -835,6 +860,7 @@ def parse_prompt_tokens_details(usage: Usage) -> PromptTokensDetailsResult:
 
     return PromptTokensDetailsResult(
         cache_hit_tokens=cache_hit_tokens,
+        cache_hit_audio_tokens=cached_audio_tokens,
         cache_creation_tokens=cache_creation_tokens,
         cache_creation_token_details=cache_creation_token_details,
         text_tokens=text_tokens,
@@ -918,7 +944,16 @@ def _calculate_input_cost(
     prompt_cost = float(prompt_tokens_details["text_tokens"]) * prompt_base_cost
 
     ### CACHE READ COST - Now uses tiered pricing
-    prompt_cost += float(prompt_tokens_details["cache_hit_tokens"]) * cache_read_cost
+    cache_hit_audio_tokens: Final = prompt_tokens_details["cache_hit_audio_tokens"]
+    audio_cache_read_rate: Final = _get_cost_per_unit(
+        model_info,
+        _get_service_tier_cost_key("cache_read_input_audio_token_cost", service_tier),
+        None,
+    )
+    prompt_cost += float(prompt_tokens_details["cache_hit_tokens"] - cache_hit_audio_tokens) * cache_read_cost
+    prompt_cost += float(cache_hit_audio_tokens) * (
+        audio_cache_read_rate if audio_cache_read_rate is not None else cache_read_cost
+    )
 
     ### AUDIO COST
     if prompt_tokens_details["audio_tokens"]:
@@ -1149,6 +1184,7 @@ def generic_cost_per_token(
     ### PROCESSING COST
     prompt_tokens_details = PromptTokensDetailsResult(
         cache_hit_tokens=0,
+        cache_hit_audio_tokens=0,
         cache_creation_tokens=0,
         cache_creation_token_details=None,
         text_tokens=usage.prompt_tokens,
@@ -1319,6 +1355,7 @@ class BilledTokenRates:
     input_cost_per_token: float
     output_cost_per_token: float
     cache_read_input_token_cost: float
+    cache_read_input_audio_token_cost: float
     cache_creation_input_token_cost: float
     cache_creation_input_token_cost_above_1hr: float
     output_cost_per_reasoning_token: float
@@ -1330,6 +1367,7 @@ class BilledTokenRates:
             input_cost_per_token=self.input_cost_per_token * multiplier,
             output_cost_per_token=self.output_cost_per_token * multiplier,
             cache_read_input_token_cost=self.cache_read_input_token_cost * multiplier,
+            cache_read_input_audio_token_cost=self.cache_read_input_audio_token_cost * multiplier,
             cache_creation_input_token_cost=self.cache_creation_input_token_cost * multiplier,
             cache_creation_input_token_cost_above_1hr=self.cache_creation_input_token_cost_above_1hr * multiplier,
             output_cost_per_reasoning_token=self.output_cost_per_reasoning_token * multiplier,
@@ -1353,15 +1391,16 @@ def _reasoning_token_count(usage: Usage) -> int:
     return parsed or _coerce_token_count(getattr(usage, "reasoning_tokens", 0))
 
 
-def _cache_token_counts(usage: Usage) -> tuple[int, int, CacheCreationTokenDetails | None]:
-    """(cache read tokens, cache creation tokens, cache creation details): read from prompt_tokens_details
-    first, then the private top-level counters the Usage constructor mirrors cache tokens onto for
-    providers/callers that bypass the details."""
+def _cache_token_counts(usage: Usage) -> tuple[int, int, int, CacheCreationTokenDetails | None]:
+    """(cache read tokens, cached audio tokens, cache creation tokens, cache creation details): read from
+    prompt_tokens_details first, then the private top-level counters the Usage constructor mirrors cache
+    tokens onto for providers/callers that bypass the details."""
     parsed: Final = parse_prompt_tokens_details(usage) if usage.prompt_tokens_details is not None else None
     parsed_read: Final = parsed["cache_hit_tokens"] if parsed is not None else 0
     parsed_creation: Final = parsed["cache_creation_tokens"] if parsed is not None else 0
     return (
         parsed_read or _coerce_token_count(getattr(usage, "_cache_read_input_tokens", 0)),
+        parsed["cache_hit_audio_tokens"] if parsed is not None else 0,
         parsed_creation or _coerce_token_count(getattr(usage, "_cache_creation_input_tokens", 0)),
         parsed["cache_creation_token_details"] if parsed is not None else None,
     )
@@ -1372,11 +1411,13 @@ def _custom_pricing_rates(custom_cost_per_token: CostPerToken) -> BilledTokenRat
     cache rates (else the input rate) and reasoning at the output rate, as _cost_per_token_custom_pricing_helper does."""
     input_rate: Final = custom_cost_per_token["input_cost_per_token"]
     output_rate: Final = custom_cost_per_token["output_cost_per_token"]
+    cache_read_rate: Final = custom_cost_per_token.get("cache_read_input_token_cost", input_rate)
     cache_creation_rate: Final = custom_cost_per_token.get("cache_creation_input_token_cost", input_rate)
     return BilledTokenRates(
         input_cost_per_token=input_rate,
         output_cost_per_token=output_rate,
-        cache_read_input_token_cost=custom_cost_per_token.get("cache_read_input_token_cost", input_rate),
+        cache_read_input_token_cost=cache_read_rate,
+        cache_read_input_audio_token_cost=cache_read_rate,
         cache_creation_input_token_cost=cache_creation_rate,
         cache_creation_input_token_cost_above_1hr=cache_creation_rate,
         output_cost_per_reasoning_token=output_rate,
@@ -1413,6 +1454,11 @@ def _cost_map_billed_rates(
         completion_base_cost=completion_base_cost,
         current_time=billing_time,
     )
+    audio_cache_read_rate: Final = _get_cost_per_unit(
+        model_info,
+        _get_service_tier_cost_key("cache_read_input_audio_token_cost", service_tier),
+        None,
+    )
     multiplier: Final = (
         _get_regional_uplift_multiplier(model_info, data_residency)
         * get_vertex_regional_endpoint_uplift(model_info, vertex_location)
@@ -1422,6 +1468,9 @@ def _cost_map_billed_rates(
         input_cost_per_token=prompt_base_cost,
         output_cost_per_token=completion_base_cost,
         cache_read_input_token_cost=cache_read_cost_rate,
+        cache_read_input_audio_token_cost=(
+            audio_cache_read_rate if audio_cache_read_rate is not None else cache_read_cost_rate
+        ),
         cache_creation_input_token_cost=cache_creation_cost_rate,
         cache_creation_input_token_cost_above_1hr=cache_creation_cost_above_1hr_rate,
         output_cost_per_reasoning_token=reasoning_rate,
@@ -1494,7 +1543,9 @@ def get_token_type_cost_breakdown(
     if rates is None:
         return TokenTypeCostBreakdown(0.0, 0.0, 0.0)
 
-    cache_read_tokens, cache_creation_tokens, cache_creation_token_details = _cache_token_counts(usage)
+    cache_read_tokens, cached_audio_tokens, cache_creation_tokens, cache_creation_token_details = _cache_token_counts(
+        usage
+    )
     cache_creation_cost: Final = (
         float(cache_creation_tokens) * rates.cache_creation_input_token_cost
         if custom_cost_per_token is not None
@@ -1507,7 +1558,10 @@ def get_token_type_cost_breakdown(
     )
     return TokenTypeCostBreakdown(
         reasoning_cost=float(_reasoning_token_count(usage)) * rates.output_cost_per_reasoning_token,
-        cache_read_cost=float(cache_read_tokens) * rates.cache_read_input_token_cost,
+        cache_read_cost=(
+            float(cache_read_tokens - cached_audio_tokens) * rates.cache_read_input_token_cost
+            + float(cached_audio_tokens) * rates.cache_read_input_audio_token_cost
+        ),
         cache_creation_cost=cache_creation_cost,
         rates=rates,
     )

litellm/llms/bedrock/realtime/handler.py

@@ -7,13 +7,20 @@ This uses aws_sdk_bedrock_runtime for bidirectional streaming with Nova Sonic.
 import asyncio
 import contextlib
 import json
-from collections.abc import AsyncIterator, Mapping
-from typing import Final, Protocol
+from collections.abc import AsyncIterator, Mapping, MutableMapping
+from dataclasses import dataclass
+from types import MappingProxyType
+from typing import Final, NoReturn, Protocol
 
 from pydantic import JsonValue, TypeAdapter
 
 import litellm
 from litellm._logging import _redact_string, verbose_proxy_logger
+from litellm.constants import (
+    BEDROCK_REALTIME_COMMITTED_FAILURE_SCOPE_KEY,
+    BEDROCK_REALTIME_PENDING_SESSION_UPDATE_SCOPE_KEY,
+    BEDROCK_REALTIME_SESSION_COMMITTED_SCOPE_KEY,
+)
 from litellm.litellm_core_utils.aws_partition import get_aws_dns_suffix
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging
 from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER
@@ -28,6 +35,32 @@ from .transformation import BedrockRealtimeConfig
 _CLIENT_MODALITIES_ADAPTER: Final[TypeAdapter["list[str] | None"]] = TypeAdapter(list[str] | None)
 _CLIENT_MESSAGE_ADAPTER: Final[TypeAdapter[JsonValue]] = TypeAdapter(JsonValue)
 
+_EMPTY_JSON_OBJECT: Final[Mapping[str, JsonValue]] = MappingProxyType({})
+
+_BEDROCK_STREAM_ERROR_STATUS: Final[Mapping[str, int]] = MappingProxyType(
+    {
+        "AccessDeniedException": 403,
+        "ConflictException": 400,
+        "InternalServerException": 500,
+        "ModelErrorException": 424,
+        "ModelNotReadyException": 429,
+        "ModelStreamErrorException": 424,
+        "ModelTimeoutException": 408,
+        "ResourceNotFoundException": 404,
+        "ServiceQuotaExceededException": 400,
+        "ServiceUnavailableException": 503,
+        "ThrottlingException": 429,
+        "ValidationException": 400,
+    }
+)
+
+
+def _as_bedrock_error(error: BaseException) -> BaseException:
+    status_code: Final = _BEDROCK_STREAM_ERROR_STATUS.get(type(error).__name__)
+    if status_code is None:
+        return error
+    return BedrockError(status_code=status_code, message=f"{type(error).__name__}: {error}")
+
 
 def _json_dict(value: JsonValue) -> dict[str, JsonValue]:
     return value if isinstance(value, dict) else {}
@@ -51,6 +84,8 @@ def _should_log_event(openai_message: Mapping[str, object]) -> bool:
 class RealtimeClientWebSocket(Protocol):
     """The client-facing websocket surface the realtime bridge talks to."""
 
+    scope: MutableMapping[str, object]  # mutable-ok: the ASGI scope is the per-connection state store
+
     async def receive_text(self) -> str: ...
 
     async def send_text(self, data: str) -> None: ...
@@ -85,6 +120,81 @@ class BedrockBidirectionalStream(Protocol):
     async def await_output(self) -> tuple[object, BedrockOutputStream]: ...
 
 
+@dataclass(frozen=True, slots=True)
+class _BridgeOutcome:
+    logged_events: tuple[OpenAIRealtimeEvents, ...]
+    provider_failure: BaseException | None
+    client_disconnected: bool
+
+
+async def _client_messages(client_ws: RealtimeClientWebSocket, initial_message: str | None) -> AsyncIterator[str]:
+    if initial_message is not None:
+        yield initial_message
+    while True:
+        try:
+            yield await client_ws.receive_text()
+        except Exception as e:  # noqa: BLE001  # any receive failure means the client is gone
+            verbose_proxy_logger.debug("Client to Bedrock forwarding ended: %s", e, exc_info=True)
+            return
+
+
+def _pending_session_update(scope: Mapping[str, object]) -> str | None:
+    """A fallback attempt on the same websocket replays the session.update the failed attempt never acked."""
+    if scope.get(BEDROCK_REALTIME_SESSION_COMMITTED_SCOPE_KEY) is True:
+        committed_failure: Final = scope.get(BEDROCK_REALTIME_COMMITTED_FAILURE_SCOPE_KEY)
+        raise BedrockError(
+            status_code=400,
+            message=(
+                "Bedrock realtime session already committed to a provider stream; it cannot be replayed"
+                + (f". The committed stream failed with: {committed_failure}" if committed_failure else "")
+            ),
+        )
+    pending: Final = scope.get(BEDROCK_REALTIME_PENDING_SESSION_UPDATE_SCOPE_KEY)
+    return pending if isinstance(pending, str) else None
+
+
+def _raise_provider_failure(scope: MutableMapping[str, object], failure: BaseException) -> NoReturn:
+    error: Final = _as_bedrock_error(failure)
+    verbose_proxy_logger.error("Bedrock Realtime: provider stream failed: %s", _redact_string(str(error)))
+    if scope.get(BEDROCK_REALTIME_SESSION_COMMITTED_SCOPE_KEY) is True:
+        scope[BEDROCK_REALTIME_COMMITTED_FAILURE_SCOPE_KEY] = _redact_string(str(error))
+    raise error from failure
+
+
+def _parse_client_message(message: str) -> Mapping[str, JsonValue]:
+    try:
+        return _json_dict(_CLIENT_MESSAGE_ADAPTER.validate_json(message))
+    except ValueError:
+        return _EMPTY_JSON_OBJECT
+
+
+async def _ack_session_update(
+    client_ws: RealtimeClientWebSocket,
+    bedrock_stream: BedrockBidirectionalStream,
+    transformation_config: BedrockRealtimeConfig,
+    model: str,
+    logging_obj: LiteLLMLogging | None,
+    parsed_client_message: Mapping[str, JsonValue],
+) -> bool:
+    """Ack the client's session.update once Bedrock accepted the stream; False means the client is gone."""
+    await bedrock_stream.await_output()
+    client_ws.scope.pop(BEDROCK_REALTIME_PENDING_SESSION_UPDATE_SCOPE_KEY, None)
+    client_ws.scope[BEDROCK_REALTIME_SESSION_COMMITTED_SCOPE_KEY] = True  # rebind-ok: scope outlives the attempt
+    if logging_obj is None:
+        return True
+    requested_modalities: Final = _CLIENT_MODALITIES_ADAPTER.validate_python(
+        _json_dict(parsed_client_message.get("session")).get("modalities")
+    )
+    try:
+        await client_ws.send_text(
+            json.dumps(transformation_config.session_updated_event(model, logging_obj, requested_modalities))
+        )
+    except Exception as e:  # noqa: BLE001  # any send failure means the client is gone
+        verbose_proxy_logger.debug("Client to Bedrock forwarding ended: %s", e, exc_info=True)
+        return False
+    return True
+
+
 class BedrockRealtime(BaseAWSLLM):
     """Handler for Bedrock Nova Sonic realtime speech-to-speech API."""
 
@@ -132,6 +242,8 @@ class BedrockRealtime(BaseAWSLLM):
         except ImportError:
             raise ImportError("Missing aws_sdk_bedrock_runtime. Install with: pip install aws-sdk-bedrock-runtime")
 
+        pending_session_update: Final = _pending_session_update(websocket.scope)
+
         # Get AWS region
         if aws_region_name is None:
             optional_params: Final = {
@@ -190,90 +302,105 @@ class BedrockRealtime(BaseAWSLLM):
 
         transformation_config: Final = BedrockRealtimeConfig()
 
-        try:
-            # Initialize the bidirectional stream
-            bedrock_stream: Final = await open_bidirectional_stream()
+        bedrock_stream: Final = await open_bidirectional_stream()
 
-            verbose_proxy_logger.debug("Bedrock Realtime: Bidirectional stream established")
+        verbose_proxy_logger.debug("Bedrock Realtime: Bidirectional stream established")
 
+        if pending_session_update is None:
             await websocket.send_text(json.dumps(transformation_config.session_created_event(model, logging_obj)))
             verbose_proxy_logger.debug("Bedrock Realtime: sent session.created to client on connect")
 
-            # Track state for transformation
-            session_state: Final[RealtimeResponseTransformInput] = {
-                "current_output_item_id": None,
-                "current_response_id": None,
-                "current_conversation_id": None,
-                "current_delta_chunks": None,
-                "current_item_chunks": None,
-                "current_delta_type": None,
-                "session_configuration_request": None,
-            }
+        # Track state for transformation
+        session_state: Final[RealtimeResponseTransformInput] = {
+            "current_output_item_id": None,
+            "current_response_id": None,
+            "current_conversation_id": None,
+            "current_delta_chunks": None,
+            "current_item_chunks": None,
+            "current_delta_type": None,
+            "session_configuration_request": None,
+        }
 
-            # Create tasks for bidirectional forwarding
-            client_to_bedrock_task: Final = asyncio.create_task(
-                self._forward_client_to_bedrock(
-                    websocket,
-                    bedrock_stream,
-                    transformation_config,
-                    model,
-                    session_state,
-                    logging_obj,
+        outcome: Final = await self._bridge(
+            websocket,
+            bedrock_stream,
+            transformation_config,
+            model,
+            session_state,
+            logging_obj,
+            initial_message=pending_session_update,
+        )
+
+        logged_events: Final = (
+            *outcome.logged_events,
+            *(
+                leftover_event
+                for leftover_event in transformation_config.leftover_usage_done_events()
+                if _should_log_event(leftover_event)
+            ),
+        )
+        if logged_events:
+            GLOBAL_LOGGING_WORKER.ensure_initialized_and_enqueue(
+                logging_obj.dispatch_success_handlers(
+                    list(logged_events),  # mutable-ok: realtime spend logging requires a list result
+                    prefer_async_handlers=True,
                 )
             )
 
-            async def forward_bedrock_and_collect_logged_events() -> tuple[OpenAIRealtimeEvents, ...]:
-                return tuple(
-                    [
-                        event
-                        async for event in self._forward_bedrock_to_client(
-                            bedrock_stream,
-                            websocket,
-                            transformation_config,
-                            model,
-                            logging_obj,
-                            session_state,
-                        )
-                    ]
-                )
-
-            bedrock_to_client_task: Final = asyncio.create_task(forward_bedrock_and_collect_logged_events())
-
-            # Wait for both tasks to complete
-            await asyncio.gather(
-                client_to_bedrock_task,
-                bedrock_to_client_task,
-                return_exceptions=True,
+        if outcome.provider_failure is None:
+            return
+        if outcome.client_disconnected:
+            verbose_proxy_logger.debug(
+                "Bedrock Realtime: stream failed after the client disconnected: %s", outcome.provider_failure
             )
+            return
+        _raise_provider_failure(websocket.scope, outcome.provider_failure)
 
-            forwarded_logged_events: Final = (
-                bedrock_to_client_task.result()
-                if not bedrock_to_client_task.cancelled() and bedrock_to_client_task.exception() is None
-                else ()
-            )
-            logged_events: Final = (
-                *forwarded_logged_events,
-                *(
-                    leftover_event
-                    for leftover_event in transformation_config.leftover_usage_done_events()
-                    if _should_log_event(leftover_event)
-                ),
-            )
-            if logged_events:
-                GLOBAL_LOGGING_WORKER.ensure_initialized_and_enqueue(
-                    logging_obj.dispatch_success_handlers(
-                        list(logged_events),  # mutable-ok: realtime spend logging requires a list result
-                        prefer_async_handlers=True,
-                    )
-                )
+    async def _bridge(
+        self,
+        websocket: RealtimeClientWebSocket,
+        bedrock_stream: BedrockBidirectionalStream,
+        transformation_config: BedrockRealtimeConfig,
+        model: str,
+        session_state: RealtimeResponseTransformInput,
+        logging_obj: LiteLLMLogging,
+        initial_message: str | None,
+    ) -> _BridgeOutcome:
+        """Run both forwarding directions until the client leaves or either side fails."""
+        logged: Final[list[OpenAIRealtimeEvents]] = []  # mutable-ok: events forwarded before a failure are still spend
 
-        except Exception as e:
-            verbose_proxy_logger.exception("Error in BedrockRealtime.async_realtime: %s", e)
-            try:
-                await websocket.close(code=1011, reason=_redact_string(f"Internal error: {e}"))
-            except Exception:
-                pass
-            raise
+        async def collect_logged_events() -> None:
+            async for event in self._forward_bedrock_to_client(
+                bedrock_stream, websocket, transformation_config, model, logging_obj, session_state
+            ):
+                logged.append(event)
+
+        client_task: Final = asyncio.create_task(
+            self._forward_client_to_bedrock(
+                websocket, bedrock_stream, transformation_config, model, session_state, logging_obj, initial_message
+            )
+        )
+        bedrock_task: Final = asyncio.create_task(collect_logged_events())
+
+        await asyncio.wait((client_task, bedrock_task), return_when=asyncio.FIRST_EXCEPTION)
+        client_disconnected: Final = (
+            client_task.done() and not client_task.cancelled() and client_task.exception() is None
+        )
+        client_task.cancel()
+        bedrock_task.cancel()
+        client_outcome, bedrock_outcome = await asyncio.gather(client_task, bedrock_task, return_exceptions=True)
+
+        return _BridgeOutcome(
+            logged_events=tuple(logged),
+            provider_failure=(
+                client_outcome
+                if isinstance(client_outcome, Exception)
+                else bedrock_outcome
+                if isinstance(bedrock_outcome, Exception)
+                else None
+            ),
+            client_disconnected=client_disconnected,
+        )
 
     async def _forward_client_to_bedrock(
         self,
@@ -283,8 +410,12 @@ class BedrockRealtime(BaseAWSLLM):
         model: str,
         session_state: RealtimeResponseTransformInput,
         logging_obj: LiteLLMLogging | None = None,
-    ):
-        """Forward messages from client WebSocket to Bedrock stream."""
+        initial_message: str | None = None,
+    ) -> None:
+        """Forward messages from client WebSocket to Bedrock stream.
+
+        Returns once the client is gone; provider failures (input stream or readiness) propagate to the caller.
+        """
         from aws_sdk_bedrock_runtime.models import (
             BidirectionalInputPayloadPart,
             InvokeModelWithBidirectionalStreamInputChunk,
@@ -299,41 +430,28 @@ class BedrockRealtime(BaseAWSLLM):
             verbose_proxy_logger.debug("Bedrock Realtime: Sent to Bedrock: %s", bedrock_message[:200])
 
         try:
-            while True:
-                # Receive message from client
-                message = await client_ws.receive_text()
+            async for message in _client_messages(client_ws, initial_message):
                 verbose_proxy_logger.debug("Bedrock Realtime: Received from client: %s", message[:200])
+                parsed_client_message = _parse_client_message(message)
+                is_session_update = _json_str(parsed_client_message.get("type")) == "session.update"
+                if is_session_update:
+                    client_ws.scope[BEDROCK_REALTIME_PENDING_SESSION_UPDATE_SCOPE_KEY] = (
+                        message  # rebind-ok: scope outlives the attempt
+                    )
 
-                # Transform OpenAI format to Bedrock format
                 transformed_messages = transformation_config.transform_realtime_request(
                     message=message,
                     model=model,
                     session_configuration_request=session_state.get("session_configuration_request"),
                 )
-
-                # Send transformed messages to Bedrock
                 for bedrock_message in transformed_messages:
                     await send_to_bedrock(bedrock_message)
 
-                if logging_obj is not None:
-                    client_message_type: str | None = None
-                    requested_modalities: list[str] | None = None
-                    with contextlib.suppress(Exception):
-                        parsed_client_message = _json_dict(_CLIENT_MESSAGE_ADAPTER.validate_json(message))
-                        client_message_type = _json_str(parsed_client_message.get("type"))
-                        if client_message_type == "session.update":
-                            requested_modalities = _CLIENT_MODALITIES_ADAPTER.validate_python(
-                                _json_dict(parsed_client_message.get("session")).get("modalities")
-                            )
-                    if client_message_type == "session.update":
-                        await client_ws.send_text(
-                            json.dumps(
-                                transformation_config.session_updated_event(model, logging_obj, requested_modalities)
-                            )
-                        )
-
-        except Exception as e:
-            verbose_proxy_logger.debug("Client to Bedrock forwarding ended: %s", e, exc_info=True)
+                if is_session_update and not await _ack_session_update(
+                    client_ws, bedrock_stream, transformation_config, model, logging_obj, parsed_client_message
+                ):
+                    break
+        finally:
             for close_message in transformation_config.session_close_messages():
                 with contextlib.suppress(Exception):
                     await send_to_bedrock(close_message)
@@ -349,68 +467,71 @@ class BedrockRealtime(BaseAWSLLM):
         logging_obj: LiteLLMLogging,
         session_state: RealtimeResponseTransformInput,
     ) -> AsyncIterator[OpenAIRealtimeEvents]:
-        """Forward messages from Bedrock to the client, yielding the ones to record for spend logging."""
-        try:
-            while True:
-                # Receive from Bedrock
-                output = await bedrock_stream.await_output()
-                result = await output[1].receive()
+        """Forward messages from Bedrock to the client, yielding the ones to record for spend logging.
 
-                if result is None:
-                    verbose_proxy_logger.debug("Bedrock Realtime: Bedrock stream ended")
-                    break
+        Provider failures propagate to the caller; the client websocket is only closed on a normal stream end.
+        """
 
-                payload_bytes = result.value.bytes_ if result.value else None
-                if payload_bytes:
-                    bedrock_response = payload_bytes.decode("utf-8")
-                    verbose_proxy_logger.debug("Bedrock Realtime: Received from Bedrock: %s", bedrock_response[:200])
-
-                    # Transform Bedrock format to OpenAI format
-                    realtime_response_transform_input: RealtimeResponseTransformInput = {
-                        "current_output_item_id": session_state.get("current_output_item_id"),
-                        "current_response_id": session_state.get("current_response_id"),
-                        "current_conversation_id": session_state.get("current_conversation_id"),
-                        "current_delta_chunks": session_state.get("current_delta_chunks"),
-                        "current_item_chunks": session_state.get("current_item_chunks"),
-                        "current_delta_type": session_state.get("current_delta_type"),
-                        "session_configuration_request": session_state.get("session_configuration_request"),
-                    }
-
-                    transformed_response = transformation_config.transform_realtime_response(
-                        message=bedrock_response,
-                        model=model,
-                        logging_obj=logging_obj,
-                        realtime_response_transform_input=realtime_response_transform_input,
-                    )
-
-                    # Update session state
-                    session_state.update(
-                        {
-                            "current_output_item_id": transformed_response.get("current_output_item_id"),
-                            "current_response_id": transformed_response.get("current_response_id"),
-                            "current_conversation_id": transformed_response.get("current_conversation_id"),
-                            "current_delta_chunks": transformed_response.get("current_delta_chunks"),
-                            "current_item_chunks": transformed_response.get("current_item_chunks"),
-                            "current_delta_type": transformed_response.get("current_delta_type"),
-                            "session_configuration_request": transformed_response.get("session_configuration_request"),
-                        }
-                    )
-
-                    # Send transformed messages to client
-                    response_value = transformed_response["response"]
-                    openai_messages = response_value if isinstance(response_value, list) else (response_value,)
-                    for openai_message in openai_messages:
-                        message_json = json.dumps(openai_message)
-                        await client_ws.send_text(message_json)
-                        verbose_proxy_logger.debug("Bedrock Realtime: Sent to client: %s", message_json[:200])
-                        if _should_log_event(openai_message):
-                            yield openai_message
-
-        except Exception as e:
-            verbose_proxy_logger.debug("Bedrock to client forwarding ended: %s", e, exc_info=True)
-        finally:
-            # Close the client WebSocket
+        async def send_to_client(message_json: str) -> bool:
             try:
-                await client_ws.close()
-            except Exception:
-                pass
+                await client_ws.send_text(message_json)
+            except Exception as e:  # noqa: BLE001  # any send failure means the client is gone
+                verbose_proxy_logger.debug("Bedrock to client forwarding ended: %s", e, exc_info=True)
+                return False
+            verbose_proxy_logger.debug("Bedrock Realtime: Sent to client: %s", message_json[:200])
+            return True
+
+        output: Final = await bedrock_stream.await_output()
+        while True:
+            result = await output[1].receive()
+
+            if result is None:
+                verbose_proxy_logger.debug("Bedrock Realtime: Bedrock stream ended")
+                with contextlib.suppress(Exception):
+                    await client_ws.close()
+                return
+
+            payload_bytes = result.value.bytes_ if result.value else None
+            if payload_bytes:
+                bedrock_response = payload_bytes.decode("utf-8")
+                verbose_proxy_logger.debug("Bedrock Realtime: Received from Bedrock: %s", bedrock_response[:200])
+
+                # Transform Bedrock format to OpenAI format
+                realtime_response_transform_input: RealtimeResponseTransformInput = {
+                    "current_output_item_id": session_state.get("current_output_item_id"),
+                    "current_response_id": session_state.get("current_response_id"),
+                    "current_conversation_id": session_state.get("current_conversation_id"),
+                    "current_delta_chunks": session_state.get("current_delta_chunks"),
+                    "current_item_chunks": session_state.get("current_item_chunks"),
+                    "current_delta_type": session_state.get("current_delta_type"),
+                    "session_configuration_request": session_state.get("session_configuration_request"),
+                }
+
+                transformed_response = transformation_config.transform_realtime_response(
+                    message=bedrock_response,
+                    model=model,
+                    logging_obj=logging_obj,
+                    realtime_response_transform_input=realtime_response_transform_input,
+                )
+
+                # Update session state
+                session_state.update(
+                    {
+                        "current_output_item_id": transformed_response.get("current_output_item_id"),
+                        "current_response_id": transformed_response.get("current_response_id"),
+                        "current_conversation_id": transformed_response.get("current_conversation_id"),
+                        "current_delta_chunks": transformed_response.get("current_delta_chunks"),
+                        "current_item_chunks": transformed_response.get("current_item_chunks"),
+                        "current_delta_type": transformed_response.get("current_delta_type"),
+                        "session_configuration_request": transformed_response.get("session_configuration_request"),
+                    }
+                )
+
+                # Send transformed messages to client
+                response_value = transformed_response["response"]
+                openai_messages = response_value if isinstance(response_value, list) else (response_value,)
+                for openai_message in openai_messages:
+                    if not await send_to_client(json.dumps(openai_message)):
+                        return
+                    if _should_log_event(openai_message):
+                        yield openai_message

litellm/model_prices_and_context_window_backup.json

@@ -5513,7 +5513,8 @@
     },
     "azure/gpt-realtime-2025-08-28": {
         "cache_creation_input_audio_token_cost": 4e-06,
-        "cache_read_input_token_cost": 4e-06,
+        "cache_read_input_audio_token_cost": 4e-07,
+        "cache_read_input_token_cost": 4e-07,
         "deprecation_date": "2027-03-02",
         "input_cost_per_audio_token": 3.2e-05,
         "input_cost_per_image_token": 5e-06,
@@ -5546,7 +5547,8 @@
     },
     "azure/gpt-realtime-1.5-2026-02-23": {
         "cache_creation_input_audio_token_cost": 4e-06,
-        "cache_read_input_token_cost": 4e-06,
+        "cache_read_input_audio_token_cost": 4e-07,
+        "cache_read_input_token_cost": 4e-07,
         "deprecation_date": "2027-08-24",
         "input_cost_per_audio_token": 3.2e-05,
         "input_cost_per_image_token": 5e-06,
@@ -5683,6 +5685,7 @@
     },
     "azure/gpt-realtime-mini": {
         "cache_creation_input_audio_token_cost": 3e-07,
+        "cache_read_input_audio_token_cost": 3e-07,
         "cache_read_input_token_cost": 6e-08,
         "input_cost_per_audio_token": 1e-05,
         "input_cost_per_image_token": 8e-07,
@@ -5715,6 +5718,7 @@
     },
     "azure/gpt-realtime-mini-2025-10-06": {
         "cache_creation_input_audio_token_cost": 3e-07,
+        "cache_read_input_audio_token_cost": 3e-07,
         "cache_read_input_token_cost": 6e-08,
         "input_cost_per_audio_token": 1e-05,
         "input_cost_per_image_token": 8e-07,
@@ -7409,6 +7413,80 @@
         "supports_web_search": true,
         "supports_xhigh_reasoning_effort": true
     },
+    "azure/gpt-chat-latest": {
+        "cache_read_input_token_cost": 5e-07,
+        "deprecation_date": "2026-12-02",
+        "input_cost_per_token": 5e-06,
+        "litellm_provider": "azure",
+        "max_input_tokens": 272000,
+        "max_output_tokens": 128000,
+        "max_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 3e-05,
+        "reasoning_effort_levels": [
+            "medium"
+        ],
+        "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/",
+        "supported_endpoints": [
+            "/v1/chat/completions",
+            "/v1/responses"
+        ],
+        "supported_modalities": [
+            "text",
+            "image"
+        ],
+        "supported_output_modalities": [
+            "text"
+        ],
+        "supports_function_calling": true,
+        "supports_native_streaming": true,
+        "supports_parallel_function_calling": true,
+        "supports_pdf_input": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "supports_vision": true,
+        "supports_web_search": true
+    },
+    "azure/chat-latest": {
+        "cache_read_input_token_cost": 5e-07,
+        "deprecation_date": "2026-12-02",
+        "input_cost_per_token": 5e-06,
+        "litellm_provider": "azure",
+        "max_input_tokens": 272000,
+        "max_output_tokens": 128000,
+        "max_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 3e-05,
+        "reasoning_effort_levels": [
+            "medium"
+        ],
+        "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/",
+        "supported_endpoints": [
+            "/v1/chat/completions",
+            "/v1/responses"
+        ],
+        "supported_modalities": [
+            "text",
+            "image"
+        ],
+        "supported_output_modalities": [
+            "text"
+        ],
+        "supports_function_calling": true,
+        "supports_native_streaming": true,
+        "supports_parallel_function_calling": true,
+        "supports_pdf_input": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "supports_vision": true,
+        "supports_web_search": true
+    },
     "azure/us/gpt-5.6": {
         "cache_creation_input_token_cost": 6.875e-06,
         "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05,
@@ -7675,6 +7753,43 @@
         "supports_web_search": true,
         "supports_xhigh_reasoning_effort": true
     },
+    "azure/us/gpt-chat-latest": {
+        "cache_read_input_token_cost": 5.5e-07,
+        "deprecation_date": "2026-12-02",
+        "input_cost_per_token": 5.5e-06,
+        "litellm_provider": "azure",
+        "max_input_tokens": 272000,
+        "max_output_tokens": 128000,
+        "max_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 3.3e-05,
+        "reasoning_effort_levels": [
+            "medium"
+        ],
+        "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/",
+        "supported_endpoints": [
+            "/v1/chat/completions",
+            "/v1/responses"
+        ],
+        "supported_modalities": [
+            "text",
+            "image"
+        ],
+        "supported_output_modalities": [
+            "text"
+        ],
+        "supports_function_calling": true,
+        "supports_native_streaming": true,
+        "supports_parallel_function_calling": true,
+        "supports_pdf_input": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "supports_vision": true,
+        "supports_web_search": true
+    },
     "azure/eu/gpt-5.6": {
         "cache_creation_input_token_cost": 6.875e-06,
         "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05,
@@ -23113,34 +23228,6 @@
         "output_cost_per_token": 0.0,
         "source": "https://fireworks.ai/pricing"
     },
-    "friendliai/meta-llama-3.1-70b-instruct": {
-        "input_cost_per_token": 6e-07,
-        "litellm_provider": "friendliai",
-        "max_input_tokens": 8192,
-        "max_output_tokens": 8192,
-        "max_tokens": 8192,
-        "mode": "chat",
-        "output_cost_per_token": 6e-07,
-        "supports_function_calling": true,
-        "supports_parallel_function_calling": true,
-        "supports_response_schema": true,
-        "supports_system_messages": true,
-        "supports_tool_choice": true
-    },
-    "friendliai/meta-llama-3.1-8b-instruct": {
-        "input_cost_per_token": 1e-07,
-        "litellm_provider": "friendliai",
-        "max_input_tokens": 8192,
-        "max_output_tokens": 8192,
-        "max_tokens": 8192,
-        "mode": "chat",
-        "output_cost_per_token": 1e-07,
-        "supports_function_calling": true,
-        "supports_parallel_function_calling": true,
-        "supports_response_schema": true,
-        "supports_system_messages": true,
-        "supports_tool_choice": true
-    },
     "friendliai/zai-org/GLM-5.3-Flash": {
         "litellm_provider": "friendliai",
         "max_input_tokens": 1048576,
@@ -32938,6 +33025,7 @@
     },
     "gpt-realtime": {
         "cache_creation_input_audio_token_cost": 4e-07,
+        "cache_read_input_audio_token_cost": 4e-07,
         "cache_read_input_token_cost": 4e-07,
         "deprecation_date": "2027-01-20",
         "input_cost_per_audio_token": 3.2e-05,
@@ -32972,6 +33060,7 @@
     },
     "gpt-realtime-1.5": {
         "cache_creation_input_audio_token_cost": 4e-07,
+        "cache_read_input_audio_token_cost": 4e-07,
         "cache_read_input_token_cost": 4e-07,
         "input_cost_per_audio_token": 3.2e-05,
         "input_cost_per_image_token": 5e-06,
@@ -33146,6 +33235,7 @@
     },
     "gpt-realtime-2025-08-28": {
         "cache_creation_input_audio_token_cost": 4e-07,
+        "cache_read_input_audio_token_cost": 4e-07,
         "cache_read_input_token_cost": 4e-07,
         "deprecation_date": "2027-01-20",
         "input_cost_per_audio_token": 3.2e-05,
@@ -57947,8 +58037,9 @@
             },
             {
                 "name": "claude-adaptive-thinking",
-                "pattern": "claude-[a-z]+-(?:4[-._](?:[6-9]|[1-9]\\d)(?!\\d)|(?:[5-9]|[1-9]\\d)(?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)",
-                "description": "Claude at version 4.6 or higher, in any id shape that contains claude-<family>-: minors 4.6 through 4.99, any later major-minor, and bare 5+ majors so a new family shaped like claude-fable-5 matches. Requiring the claude- prefix keeps non-Claude names such as team-sonnet-5-1 out. The minor is capped at two digits so an 8-digit date suffix such as claude-opus-4-20250514 is never read as a >= 4.6 minor. Turns on adaptive thinking for new versions and new families with no code change.",
+                "pattern": "claude-[a-z]+-(?:4[-._](?:[6-9]|[1-9]\\d)(?!\\d)|[5-9](?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)",
+                "fill_missing_for_providers": ["anthropic", "azure_ai", "bedrock", "bedrock_converse", "vertex_ai-anthropic_models"],
+                "description": "Claude at version 4.6 or higher, in any id shape that contains claude-<family>-: minors 4.6 through 4.99, any later major-minor, and bare 5+ majors so a new family shaped like claude-fable-5 matches. Requiring the claude- prefix keeps non-Claude names such as team-sonnet-5-1 out. The minor is capped at two digits so an 8-digit date suffix such as claude-opus-4-20250514 is never read as a >= 4.6 minor. Two-digit majors are deliberately not matched so ids like claude-opus-41 (4.1) are not read as major 41. Turns on adaptive thinking for new versions and new families with no code change.",
                 "model_info": {
                     "supports_adaptive_thinking": true
                 }
@@ -57956,6 +58047,7 @@
             {
                 "name": "claude-legacy-thinking",
                 "pattern": "claude-[a-z]+-4[-._]6(?!\\d)",
+                "fill_missing_for_providers": ["anthropic", "azure_ai", "bedrock", "bedrock_converse", "vertex_ai-anthropic_models"],
                 "description": "Claude at version 4.6 exactly, in any id shape that contains claude-<family>-4-6 (dotted and underscored minors included, dated releases such as claude-sonnet-4-6-20260219 too). The 4.6 family is adaptive-thinking yet still accepts legacy thinking.type=enabled with budget_tokens, so the caller's hard budget cap is forwarded verbatim instead of being rewritten to an uncapped output_config.effort. The lookahead keeps two-digit minors such as 4-60 from matching. 4.7+ and 5+ majors reject the legacy shape and stay on the adaptive translation.",
                 "model_info": {
                     "supports_legacy_thinking": true
@@ -57971,8 +58063,9 @@
             },
             {
                 "name": "claude-mid-conversation-system",
-                "pattern": "claude-[a-z]+-(?:4[-._](?:[89]|[1-9]\\d)(?!\\d)|(?:[5-9]|[1-9]\\d)(?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)",
-                "description": "Claude at version 4.8 or higher, in any id shape that contains claude-<family>-: minors 4.8 through 4.99, any later major-minor, and bare 5+ majors so a new family like claude-fable-5 matches. Anthropic introduced mid-conversation system messages with Opus 4.8 and every newer Claude keeps them; 4.7 and below reject the system role inside messages.",
+                "pattern": "claude-[a-z]+-(?:4[-._](?:[89]|[1-9]\\d)(?!\\d)|[5-9](?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)",
+                "fill_missing_for_providers": ["anthropic", "azure_ai", "bedrock", "bedrock_converse", "vertex_ai-anthropic_models"],
+                "description": "Claude at version 4.8 or higher, in any id shape that contains claude-<family>-: minors 4.8 through 4.99, any later major-minor, and bare 5+ majors so a new family like claude-fable-5 matches. Two-digit majors are deliberately not matched so ids like claude-opus-41 (4.1) are not read as major 41. Anthropic introduced mid-conversation system messages with Opus 4.8 and every newer Claude keeps them; 4.7 and below reject the system role inside messages.",
                 "model_info": {
                     "supports_mid_conversation_system": true
                 }
@@ -57988,6 +58081,7 @@
             {
                 "name": "openai-reasoning-family-baseline",
                 "pattern": "^(?!.*search-api)(?:[a-z0-9_.-]+/)*(?:ft:)?(?:o[1-9]\\d*(?![a-z0-9])|gpt-[5-9](?:\\.\\d+)?(?![0-9.])|(?:gpt-\\d+(?:\\.\\d+)?(?:-[a-z0-9]+)*-)?(?:codex|deep-research|chat-latest)(?![a-z0-9]))",
+                "fill_missing_for_providers": ["azure", "azure_ai", "openai"],
                 "description": "OpenAI reasoning families by id shape, under any provider namespace and with an optional ft: prefix: the o-series (o1, o3-pro, o4-mini), gpt-5 through gpt-9 majors including dotted minors and suffixed variants (gpt-5.5-cyber, gpt-6-astra), and the codex, deep-research and chat-latest lines when standalone or on a gpt base. gpt-5-search-api is excluded because it is a search-only surface. Every model here is a reasoning model, and the Responses API drops the caller's reasoning param for any mapped OpenAI model whose info lacks supports_reasoning, so an id the registry has not named yet keeps its reasoning settings instead of silently losing them. Rules lose to exact entries. Carries no mode and no pricing, so cost stays on the standard unpriced behavior.",
                 "model_info": {
                     "supports_reasoning": true

litellm/proxy/_types.py

@@ -2635,9 +2635,13 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase):
         None,
         description="max file size in MB for /v1/files uploads, for any purpose, if a file is larger than this size it will be rejected before being forwarded to the provider",
     )
+    allowed_file_extensions: tuple[str, ...] | None = Field(
+        None,
+        description="the only file extensions (e.g. ['.jsonl', '.pdf', '.txt']) accepted on /v1/files uploads, for any purpose, matched case-insensitively against the uploaded filename. Files with any other extension, or none, are rejected. An empty list rejects every upload. Unset means no allowlist is applied",
+    )
     blocked_file_extensions: tuple[str, ...] | None = Field(
         None,
-        description="file extensions (e.g. ['.exe', '.sh']) rejected on /v1/files uploads, for any purpose, matched case-insensitively against the uploaded filename",
+        description="file extensions (e.g. ['.exe', '.sh']) rejected on /v1/files uploads, for any purpose, matched case-insensitively against the uploaded filename. Deprecated in favour of allowed_file_extensions; still enforced, after the allowlist, when set",
     )
     max_response_size_mb: int | None = Field(
         None,
@@ -4837,6 +4841,14 @@ class JWTIssuerConfig(BaseModel):
         default=None,
         description="Issuer-specific claim path to normalize into LiteLLM's end-user id.",
     )
+    virtual_key_claim_field: str | None = Field(
+        default=None,
+        description="Issuer-specific claim path used for the virtual key mapping lookup. Falls back to the global field.",
+    )
+    unregistered_jwt_client_behavior: UnregisteredJWTClientBehavior | None = Field(
+        default=None,
+        description="Issuer-specific policy when the virtual key claim has no mapping. Falls back to the global policy.",
+    )
 
     model_config = {
         "extra": "forbid",
@@ -5063,6 +5075,28 @@ class LiteLLM_JWTAuth(LiteLLMPydanticObjectBase):
 
         super().__init__(**kwargs)
 
+    def get_issuer_config(self, issuer: str | None) -> JWTIssuerConfig | None:
+        if issuer is None or self.issuers is None:
+            return None
+        return next((config for config in self.issuers if config.issuer == issuer), None)
+
+    def is_virtual_key_mapping_configured(self) -> bool:
+        if self.virtual_key_claim_field is not None:
+            return True
+        return any(config.virtual_key_claim_field is not None for config in self.issuers or ())
+
+    def get_virtual_key_claim_field(self, issuer: str | None) -> str | None:
+        issuer_config: Final = self.get_issuer_config(issuer)
+        if issuer_config is not None and issuer_config.virtual_key_claim_field is not None:
+            return issuer_config.virtual_key_claim_field
+        return self.virtual_key_claim_field
+
+    def get_unregistered_jwt_client_behavior(self, issuer: str | None) -> UnregisteredJWTClientBehavior:
+        issuer_config: Final = self.get_issuer_config(issuer)
+        if issuer_config is not None and issuer_config.unregistered_jwt_client_behavior is not None:
+            return issuer_config.unregistered_jwt_client_behavior
+        return self.unregistered_jwt_client_behavior
+
 
 class PrismaCompatibleUpdateDBModel(TypedDict, total=False):
     model_name: str

litellm/proxy/auth/auth_checks.py

@@ -123,6 +123,7 @@ from litellm.repositories.table_repositories import (
 from litellm.repositories.team_repository import TeamRepository
 from litellm.repositories.user_repository import UserRepository
 from litellm.router import Router
+from litellm.types.proxy.auth.auth_checks import UserNotFoundError
 from litellm.types.proxy.model_access_group_budget import ModelAccessGroupBudget
 from litellm.utils import get_utc_datetime
 
@@ -2375,13 +2376,6 @@ async def _backfill_null_user_email(
     return updated_row
 
 
-class UserNotFoundError(ValueError):
-    """The user row is provably absent, as opposed to merely unreadable, so a caller that reads a missing row as no user-level limits can key on it without also swallowing a database that would not answer."""
-
-    def __init__(self, user_id: str) -> None:
-        super().__init__(f"User doesn't exist in db. 'user_id'={user_id}. Create user via `/user/new` call.")
-
-
 @log_db_metrics
 async def get_user_object(
     user_id: str | None,

litellm/proxy/auth/resolvers/grants.py

@@ -28,11 +28,11 @@ from litellm.proxy._types import (
 )
 from litellm.proxy.auth.auth_checks import (
     TeamNotFoundError,
-    UserNotFoundError,
     get_team_membership,
     get_team_object,
     get_user_object,
 )
+from litellm.types.proxy.auth.auth_checks import UserNotFoundError
 
 if TYPE_CHECKING:
     from litellm.proxy._types import Span

litellm/proxy/auth/route_checks.py

@@ -136,6 +136,9 @@ class RouteChecks:
                     #  For llm_api_routes, also check registered pass-through endpoints
                     ################################################
                     if allowed_route == "llm_api_routes":
+                        if route == "/auto_router/session" and RouteChecks._get_request_method(request) == "GET":
+                            return True
+
                         from litellm.proxy.pass_through_endpoints.pass_through_endpoints import (
                             InitPassThroughEndpointHelpers,
                         )

litellm/proxy/auth/user_api_key_auth.py

@@ -996,9 +996,12 @@ async def _resolve_jwt_to_virtual_key(
       - Raises HTTPException: REJECT policy hit, missing claim under
         REJECT/AUTO_REGISTER, or other policy violations.
     """
-    virtual_key_claim_field: Final = jwt_handler.litellm_jwtauth.virtual_key_claim_field
+    raw_issuer: Final = jwt_claims.get(JWTHandler.LITELLM_JWT_ISSUER_CLAIM)
+    normalized_issuer: Final = raw_issuer if isinstance(raw_issuer, str) else None
+    virtual_key_claim_field: Final = jwt_handler.litellm_jwtauth.get_virtual_key_claim_field(normalized_issuer)
     if virtual_key_claim_field is None:
         return None
+    behavior: Final = jwt_handler.litellm_jwtauth.get_unregistered_jwt_client_behavior(normalized_issuer)
 
     claim_value: Final = get_nested_value(
         data=jwt_claims,
@@ -1015,7 +1018,6 @@ async def _resolve_jwt_to_virtual_key(
         # simply by presenting a JWT that omits the configured field. For
         # AUTO_REGISTER there is no stable identity to map without a claim
         # value, so we deny rather than create a sentinel-keyed record.
-        behavior = jwt_handler.litellm_jwtauth.unregistered_jwt_client_behavior
         if behavior in (
             UnregisteredJWTClientBehavior.REJECT,
             UnregisteredJWTClientBehavior.AUTO_REGISTER,
@@ -1030,7 +1032,13 @@ async def _resolve_jwt_to_virtual_key(
         return None
 
     cache_key: Final = jwt_key_mapping_cache_key(virtual_key_claim_field, str(claim_value))
-    cached_mapping: Final = await user_api_key_cache.async_get_cache(cache_key)
+    raw_cached_mapping: Final = await user_api_key_cache.async_get_cache(cache_key)
+    sentinel_written_by_this_policy: Final = behavior == UnregisteredJWTClientBehavior.AUTO_REGISTER
+    cached_mapping: Final = (
+        None
+        if raw_cached_mapping == _JWT_PROXY_ADMIN_SENTINEL and not sentinel_written_by_this_policy
+        else raw_cached_mapping
+    )
 
     if cached_mapping == _JWT_PROXY_ADMIN_SENTINEL:
         # Previously resolved to a proxy admin via auth_builder; skip the
@@ -1039,7 +1047,6 @@ async def _resolve_jwt_to_virtual_key(
         return None
 
     if cached_mapping == "__NO_MAPPING__":
-        behavior = jwt_handler.litellm_jwtauth.unregistered_jwt_client_behavior
         if behavior == UnregisteredJWTClientBehavior.REJECT:
             raise HTTPException(
                 status_code=403,
@@ -1102,8 +1109,6 @@ async def _resolve_jwt_to_virtual_key(
         )
 
     # No mapping found (DB miss or no DB) — apply no-match policy.
-    behavior = jwt_handler.litellm_jwtauth.unregistered_jwt_client_behavior
-
     if behavior == UnregisteredJWTClientBehavior.REJECT:
         # Cache the miss before raising so repeated rejections are served from
         # cache and don't re-query the DB on every request.
@@ -1483,7 +1488,7 @@ async def _user_api_key_auth_builder(
                 # unnecessary DB queries in auth_builder
                 do_standard_jwt_auth = True
                 pending_auto_register: _PendingAutoRegister | None = None
-                if jwt_handler.litellm_jwtauth.virtual_key_claim_field is not None:
+                if jwt_handler.litellm_jwtauth.is_virtual_key_mapping_configured():
                     # Decode JWT to get claims without running full auth_builder
                     jwt_claims: dict | None
                     if jwt_handler.litellm_jwtauth.oidc_userinfo_enabled and not is_jwt:

litellm/proxy/client/cli/README.md

@@ -585,7 +585,9 @@ LiteLLM       ████████░░░░░░░░░░░░░░
 Claude Opus 5 ████████████████████████  $0.38
 ```
 
-The routed model comes from Claude Code's own transcript, so it only names the tier model when the auto-router deployment sets `return_raw_model_name: true` (the `lite autoroute` wizard does); otherwise it shows the alias you requested. The cost lines come from `GET /auto_router/session?session_id=...`, which any virtual key may call for its own sessions, and are cached for five seconds under a per-user `$TMPDIR/litellm-statusline-<uid>` directory. The baseline is the priciest model in the router's hardest tier, the same counterfactual the auto-router's savings reports use. `lite unconfigure claude` removes the `statusLine` entry only while it still points at that script.
+After the first response, the status line uses the latest routed model recorded by `GET /auto_router/session?session_id=...`, so it can show the tier model even when the transcript contains the router alias. If no session record is available, it falls back to Claude Code's transcript. Session records and costs are cached for five seconds under a per-user `$TMPDIR/litellm-statusline-<uid>` directory. The gateway records turns asynchronously, so the display can briefly lag a completed turn. Any virtual key may read its own sessions. The baseline is the priciest model in the router's hardest tier, the same counterfactual the auto-router's savings reports use. `lite unconfigure claude` removes the `statusLine` entry only while it still points at that script
+
+After upgrading the CLI, rerun your original `lite configure claude` command with the same gateway, key and model choice to refresh `~/.litellm/statusline.py`. Keep any explicit `--model` value: omitting it removes the earlier model pin. Package upgrades alone do not refresh this installed copy
 
 `lite codex` registers the same script as a Codex `Stop` hook for the launch, so after each turn Codex prints the same block as a system message. Codex asks once to trust the hook; the answer is remembered for later launches.
 

litellm/proxy/client/cli/commands/statusline_script.py

@@ -7,19 +7,17 @@ status refresh (about every 300ms while typing), so the proxy is asked at most o
 TTL per session and every other refresh is served from a small on-disk cache that holds
 only the proxy's answer, never the key.
 
-Claude Code pipes a JSON payload on stdin (session_id, transcript_path, model); the routed
-model is the `message.model` of the latest foreground assistant line in the transcript,
-which is the proxy's response `model` field. That only names the tier model when the
-auto-router deployment sets `return_raw_model_name: true`; otherwise it is the alias the
-client requested. Codex pipes its Stop event instead (hook_event_name, session_id) and has
-no transcript to read, so the routed model comes from the proxy's session record and the
-result is printed as a `systemMessage` for the transcript. The proxy key is read from the
-agent's own environment (the static token `lite configure claude` writes); nothing here
-spawns a credential helper.
+Claude Code pipes a JSON payload on stdin (session_id, transcript_path, model). After the
+first foreground assistant response, the routed model comes from the proxy's session
+record, falling back to the latest foreground assistant `message.model` in the transcript
+when no record is available. Codex pipes its Stop event instead (hook_event_name, session_id)
+and prints the session record as a `systemMessage` for the transcript. The proxy key is read
+from the agent's own environment (the static token `lite configure claude` writes); nothing
+here spawns a credential helper.
 
-Cost figures come from GET /auto_router/session on the proxy, which reads the per-session
-rollup written by the spend flush. That flush is asynchronous, so a turn's cost lands a
-second or two after the turn; the cache TTL absorbs it.
+The routed model and cost figures come from GET /auto_router/session on the proxy, which
+reads the per-session rollup written by the asynchronous spend flush. The record and cache
+can briefly lag a completed turn.
 """
 
 from __future__ import annotations
@@ -348,7 +346,8 @@ def status_line(
     if not session_id or not credentials.usable:
         return render(label, None, config_dir, color_enabled(env))
     session: Final = load_session(credentials, session_id, cache_dir, fetch)
-    return render(label, session, config_dir, color_enabled(env))
+    routed_label: Final = model_label(session.last_model, config_dir) if session is not None else label
+    return render(routed_label, session, config_dir, color_enabled(env))
 
 
 def codex_stop_message(

litellm/proxy/common_request_processing.py

@@ -3452,15 +3452,13 @@ class ProxyBaseLLMRequestProcessing:
             # a failed request reports no timing, matching /v1/chat/completions
             read_timing_from_logging_obj=False,
         )
-        # Extract headers from exception - check both e.headers and e.response.headers
         headers = getattr(e, "headers", None) or {}
         if not headers:
-            # Try to get headers from e.response.headers (httpx.Response)
             _response: Final = attribute_of(e, "response")
-            if _response is not None:
-                _response_headers: Final = getattr(_response, "headers", None)
-                if _response_headers:
-                    headers = get_response_headers(dict(_response_headers))
+            _response_headers: Final = getattr(_response, "headers", None) if _response is not None else None
+            _provider_headers: Final = _response_headers or getattr(e, "litellm_response_headers", None)
+            if _provider_headers:
+                headers = get_response_headers(dict(_provider_headers))
         headers.update(custom_headers)
 
         # Call response headers hook for failure

litellm/proxy/common_utils/html_forms/default_credentials_hint.py

@@ -0,0 +1,10 @@
+import os
+from collections.abc import Mapping
+
+
+def should_hide_default_credentials_hint(general_settings: Mapping[str, object]) -> bool:
+    return (
+        os.getenv("LITELLM_HIDE_DEFAULT_CREDENTIALS_HINT", "false").lower() == "true"
+        or general_settings.get("hide_default_credentials_hint", False) is True
+        or bool(os.getenv("UI_PASSWORD"))
+    )

litellm/proxy/discovery_endpoints/ui_discovery_endpoints.py

@@ -4,6 +4,7 @@ from typing import Final
 
 from fastapi import APIRouter
 
+from litellm.proxy.common_utils.html_forms.default_credentials_hint import should_hide_default_credentials_hint
 from litellm.types.proxy.discovery_endpoints.ui_discovery_endpoints import (
     UiDiscoveryEndpoints,
 )
@@ -23,10 +24,7 @@ async def get_ui_config():
         or general_settings.get("auto_redirect_ui_login_to_sso", False) is True
     )
     admin_ui_disabled: Final = os.getenv("DISABLE_ADMIN_UI", "false").lower() == "true"
-    hide_default_credentials_hint: Final = bool(
-        os.getenv("LITELLM_HIDE_DEFAULT_CREDENTIALS_HINT", "false").lower() == "true"
-        or general_settings.get("hide_default_credentials_hint", False) is True
-    )
+    hide_default_credentials_hint: Final = should_hide_default_credentials_hint(general_settings)
 
     sso_configured: Final = has_user_setup_sso()
 

litellm/proxy/hooks/rate_limiter_utils.py

@@ -6,11 +6,10 @@ from typing import Final
 
 import litellm
 from litellm._logging import verbose_proxy_logger
+from litellm.constants import PROXY_LLM_PROVIDER_FALLBACK
 from litellm.types.router import ModelGroupInfo
 from litellm.types.utils import PriorityReservationDict
 
-PROXY_LLM_PROVIDER_FALLBACK: Final = "litellm_proxy"
-
 
 def resolve_llm_provider_for_rate_limit(
     model: str | None,

litellm/proxy/management_endpoints/team_endpoints.py

@@ -156,6 +156,7 @@ from litellm.repositories.verification_token_repository import (
     VerificationTokenRepository,
 )
 from litellm.router import Router
+from litellm.types.proxy.auth.auth_checks import UserNotFoundError
 from litellm.types.proxy.management_endpoints.common_daily_activity import (
     SpendAnalyticsPaginatedResponse,
 )
@@ -179,6 +180,7 @@ from litellm.types.proxy.management_endpoints.team_endpoints import (
 if TYPE_CHECKING:
     from prisma import Prisma
     from prisma import models as prisma_models
+    from prisma import types as prisma_types
 
 router: Final = APIRouter()
 
@@ -4857,6 +4859,26 @@ async def _get_org_admin_org_ids(
     return org_ids if org_ids else None
 
 
+async def _get_user_team_ids_from_db(
+    user_id: str,
+    prisma_client: PrismaClient,
+    user_api_key_cache: UserApiKeyCache,
+    proxy_logging_obj: ProxyLogging,
+) -> tuple[str, ...]:
+    try:
+        user: Final = await get_user_object(
+            user_id=user_id,
+            prisma_client=prisma_client,
+            user_api_key_cache=user_api_key_cache,
+            user_id_upsert=False,
+            proxy_logging_obj=proxy_logging_obj,
+            check_db_only=True,
+        )
+    except UserNotFoundError:
+        return ()
+    return tuple(user.teams or ()) if user is not None else ()
+
+
 async def _build_team_list_where_conditions(
     prisma_client: PrismaClient,
     team_id: str | None,
@@ -4867,12 +4889,16 @@ async def _build_team_list_where_conditions(
     search: str | None = None,
     search_team_id_match: TeamIdSearchMatch = "exact",
     org_admin_org_ids: list[str] | None = None,
+    own_team_ids: tuple[str, ...] = (),
     user_api_key_cache: UserApiKeyCache | None = None,
     proxy_logging_obj: ProxyLogging | None = None,
 ) -> dict[str, object] | None:
     """
     Build where conditions for team list query.
 
+    An org admin listing their own teams sees the union of the teams in the
+    orgs they administer and `own_team_ids`, the teams they are a member of.
+
     Returns None when the query is guaranteed to yield no results (e.g. user
     has no team memberships), allowing the caller to skip the DB round-trip.
     """
@@ -4895,6 +4921,11 @@ async def _build_team_list_where_conditions(
 
     if organization_id:
         where_conditions["organization_id"] = organization_id
+    elif org_admin_org_ids is not None and own_team_ids:
+        org_or_membership_scope: Final[prisma_types.LiteLLM_TeamTableWhereInput] = {
+            "OR": [{"organization_id": {"in": org_admin_org_ids}}, {"team_id": {"in": list(own_team_ids)}}]
+        }
+        where_conditions["AND"] = [org_or_membership_scope]
     elif org_admin_org_ids is not None:
         # Org admin: always scope to their orgs, even when filtering by user_id.
         where_conditions["organization_id"] = {"in": org_admin_org_ids}
@@ -5026,66 +5057,72 @@ async def _enforce_list_team_v2_access(
     prisma_client: PrismaClient,
     user_api_key_cache: UserApiKeyCache,
     proxy_logging_obj: ProxyLogging,
-) -> tuple[str | None, list[str] | None]:
+) -> tuple[str | None, list[str] | None, tuple[str, ...]]:
     """Enforce access control for list_team_v2.
 
     - Proxy admins and admin viewers can query any teams.
-    - Org admins can query teams within their organizations.
+    - Org admins can query teams within their organizations, plus the teams
+      they are a member of when listing their own teams.
     - Regular users can only query their own teams.
 
-    Returns the (possibly overridden) user_id and org_admin_org_ids.
+    Returns the (possibly overridden) user_id, org_admin_org_ids and, for an
+    org admin's own query, the caller's own team ids.
     """
     is_proxy_admin: Final = _user_has_admin_view(user_api_key_dict)
-    org_admin_org_ids: list[str] | None = None
+    caller_user_id: Final = user_api_key_dict.user_id
 
     if is_proxy_admin:
-        return user_id, org_admin_org_ids
+        return user_id, None, ()
 
     # Always check org admin status so that even own-queries see
     # the full set of organisation teams, not just direct memberships.
-    if user_api_key_dict.user_id:
-        org_admin_org_ids = await _get_org_admin_org_ids(
-            user_id=user_api_key_dict.user_id,
+    org_admin_org_ids: Final = (
+        await _get_org_admin_org_ids(
+            user_id=caller_user_id,
             prisma_client=prisma_client,
             user_api_key_cache=user_api_key_cache,
             proxy_logging_obj=proxy_logging_obj,
         )
+        if caller_user_id
+        else None
+    )
 
-    if org_admin_org_ids is not None:
+    if caller_user_id and org_admin_org_ids is not None:
         # Org admin: validate org_id filter if provided
         if organization_id and organization_id not in org_admin_org_ids:
             raise HTTPException(
                 status_code=403,
                 detail={"error": "You can only view teams within your organizations."},
             )
-        # When the caller is an org admin querying their own teams (or no
-        # specific user), null out user_id so that
-        # _build_team_list_where_conditions scopes only by organization_id
-        # — org admins should see all teams in their orgs, not just teams
-        # they are a direct member of.  Keep user_id when the org admin
-        # explicitly queries a *different* user's teams.
-        if user_id is None or user_id == user_api_key_dict.user_id:
-            user_id = None
+        is_own_query: Final = user_id is None or user_id == caller_user_id
+        own_team_ids: Final = (
+            await _get_user_team_ids_from_db(
+                user_id=caller_user_id,
+                prisma_client=prisma_client,
+                user_api_key_cache=user_api_key_cache,
+                proxy_logging_obj=proxy_logging_obj,
+            )
+            if is_own_query
+            else ()
+        )
         verbose_proxy_logger.debug(
             "list_team_v2: org admin access for user=%s, org_ids=%s, user_id_filter=%s",
-            user_api_key_dict.user_id,
+            _sanitize_for_log(caller_user_id),
             org_admin_org_ids,
-            user_id,
+            _sanitize_for_log(None if is_own_query else user_id),
         )
-    else:
-        # Not an org admin — fall back to standard route check
-        if not allowed_route_check_inside_route(user_api_key_dict=user_api_key_dict, requested_user_id=user_id):
-            raise HTTPException(
-                status_code=401,
-                detail={
-                    "error": f"Only admin users can query all teams/other teams. Your user role={user_api_key_dict.user_role}"
-                },
-            )
-        # Regular user — auto-inject caller's user_id
-        if user_id is None:
-            user_id = user_api_key_dict.user_id
+        return None if is_own_query else user_id, org_admin_org_ids, own_team_ids
 
-    return user_id, org_admin_org_ids
+    # Not an org admin — fall back to standard route check
+    if not allowed_route_check_inside_route(user_api_key_dict=user_api_key_dict, requested_user_id=user_id):
+        raise HTTPException(
+            status_code=401,
+            detail={
+                "error": f"Only admin users can query all teams/other teams. Your user role={user_api_key_dict.user_role}"
+            },
+        )
+    # Regular user — auto-inject caller's user_id
+    return user_id if user_id is not None else caller_user_id, None, ()
 
 
 @router.get(
@@ -5163,7 +5200,7 @@ async def list_team_v2(
         )
 
     # --- Access control ---
-    user_id, org_admin_org_ids = await _enforce_list_team_v2_access(
+    user_id, org_admin_org_ids, own_team_ids = await _enforce_list_team_v2_access(
         user_api_key_dict=user_api_key_dict,
         user_id=user_id,
         organization_id=organization_id,
@@ -5195,6 +5232,7 @@ async def list_team_v2(
         search=search,
         search_team_id_match=search_team_id_match,
         org_admin_org_ids=org_admin_org_ids,
+        own_team_ids=own_team_ids,
         user_api_key_cache=user_api_key_cache,
         proxy_logging_obj=proxy_logging_obj,
     )
@@ -5291,17 +5329,16 @@ async def _authorize_and_filter_teams(
 
     - Proxy admins: all teams (or filtered by user_id if provided).
     - Org admins: teams from their orgs (scoped to user_id if provided).
-    - Own query (user_id matches caller): teams the user is a member of.
+    - Own query (user_id matches caller): teams the user is a member of, across all orgs.
     - Others: 401.
     """
     is_proxy_admin: Final = _user_has_admin_view(user_api_key_dict)
+    is_own_query: Final = (
+        user_id is not None and user_api_key_dict.user_id is not None and user_api_key_dict.user_id == user_id
+    )
     allowed_org_ids: list[str] | None = None
 
     if not is_proxy_admin:
-        is_own_query: Final = (
-            user_id is not None and user_api_key_dict.user_id is not None and user_api_key_dict.user_id == user_id
-        )
-
         # Check if user is an org admin (even for own queries, so they see org teams)
         if user_api_key_dict.user_id is not None:
             caller_user: Final = await get_user_object(
@@ -5328,33 +5365,30 @@ async def _authorize_and_filter_teams(
                 },
             )
 
-    if allowed_org_ids is not None:
-        # Org admin: query DB for teams in their orgs
+    if allowed_org_ids is not None and not is_own_query:
         org_teams: Final = await _raw_team_db(TeamRepository(prisma_client)).find_many(
             where={"organization_id": {"in": allowed_org_ids}},
             include={"litellm_model_table": True},
         )
         if not user_id:
             return list(org_teams)
-        # Filter org teams to only those where the target user is a member
         return [
             team
             for team in org_teams
             if team.members_with_roles and any(m.get("user_id") == user_id for m in team.members_with_roles)
         ]
-    elif user_id:
-        # Regular user: fetch all and filter by membership (Prisma can't filter JSON arrays)
-        response: Final = await _raw_team_db(TeamRepository(prisma_client)).find_many(
-            include={"litellm_model_table": True}
-        )
-        return [
-            team
-            for team in response
-            if team.members_with_roles and any(m.get("user_id") == user_id for m in team.members_with_roles)
-        ]
-    else:
+
+    response: Final = await _raw_team_db(TeamRepository(prisma_client)).find_many(include={"litellm_model_table": True})
+    if not user_id:
         # Proxy admin: all teams
-        return list(await _raw_team_db(TeamRepository(prisma_client)).find_many(include={"litellm_model_table": True}))
+        return list(response)
+
+    # Prisma can't filter JSON arrays, so membership is filtered in Python
+    return [
+        team
+        for team in response
+        if team.members_with_roles and any(m.get("user_id") == user_id for m in team.members_with_roles)
+    ]
 
 
 @router.get("/team/list", tags=["team management"], dependencies=[Depends(user_api_key_auth)])

litellm/proxy/management_endpoints/ui_sso.py

@@ -101,6 +101,7 @@ from litellm.proxy.common_utils.admin_ui_utils import (
     admin_ui_disabled,
     show_missing_vars_in_env,
 )
+from litellm.proxy.common_utils.html_forms.default_credentials_hint import should_hide_default_credentials_hint
 from litellm.proxy.common_utils.html_forms.jwt_display_template import (
     jwt_display_template,
 )
@@ -1110,10 +1111,7 @@ async def google_login(
 
     from fastapi.responses import HTMLResponse
 
-    hide_default_credentials_hint: Final = (
-        os.getenv("LITELLM_HIDE_DEFAULT_CREDENTIALS_HINT", "false").lower() == "true"
-        or general_settings.get("hide_default_credentials_hint", False) is True
-    )
+    hide_default_credentials_hint: Final = should_hide_default_credentials_hint(general_settings)
     form_response: Final = HTMLResponse(
         content=build_ui_login_form(
             show_deprecation_banner=True,

litellm/proxy/openai_files_endpoints/files_endpoints.py

@@ -78,6 +78,7 @@ from litellm.proxy.openai_files_endpoints.common_utils import (
 )
 from litellm.proxy.openai_files_endpoints.general_upload_validation import (
     MB,
+    check_allowed_extension,
     check_blocked_extension,
     check_unsafe_filename,
     check_upload_file_size,
@@ -473,6 +474,11 @@ async def create_file(
         if general_size_failure is not None:
             raise_upload_validation_failure(general_size_failure)
 
+        allowed_extensions: Final = coerce_optional_str_list_setting(general_settings.get("allowed_file_extensions"))
+        allowed_extension_failure: Final = check_allowed_extension(file.filename, allowed_extensions)
+        if allowed_extension_failure is not None:
+            raise_upload_validation_failure(allowed_extension_failure)
+
         blocked_extensions: Final = coerce_optional_str_list_setting(general_settings.get("blocked_file_extensions"))
         blocked_extension_failure: Final = check_blocked_extension(file.filename, blocked_extensions)
         if blocked_extension_failure is not None:

litellm/proxy/openai_files_endpoints/general_upload_validation.py

@@ -2,8 +2,8 @@
 Upload validation applied to every purpose at POST /v1/files.
 
 batch_file_validation.py checks the JSONL shape of purpose="batch" uploads; this
-module applies the same fast-fail-before-forwarding shape (size cap, blocked
-extensions, path-traversal filenames) regardless of purpose.
+module applies the same fast-fail-before-forwarding shape (size cap, allowed and
+blocked extensions, path-traversal filenames) regardless of purpose.
 """
 
 from dataclasses import dataclass
@@ -31,10 +31,9 @@ def coerce_optional_int_setting(raw: object) -> int | None:
     raise TypeError(f"expected an integer, got {raw!r}")
 
 
-def coerce_optional_str_list_setting(raw: object) -> tuple[str, ...]:
-    """A general_settings value declared as an optional list of strings, e.g. blocked_file_extensions."""
+def coerce_optional_str_list_setting(raw: object) -> tuple[str, ...] | None:
     if raw is None:
-        return ()
+        return None
     if not isinstance(raw, list) or not all(isinstance(item, str) for item in raw):
         raise TypeError(f"expected a list of strings, got {raw!r}")
     return tuple(raw)
@@ -46,6 +45,11 @@ class UploadedFileTooLarge:
     limit_mb: int
 
 
+@dataclass(frozen=True, slots=True)
+class UploadedFileExtensionNotAllowed:
+    extension: str
+
+
 @dataclass(frozen=True, slots=True)
 class UploadedFileBlockedExtension:
     extension: str
@@ -56,7 +60,9 @@ class UploadedFileUnsafeFilename:
     filename: str
 
 
-UploadValidationFailure = UploadedFileTooLarge | UploadedFileBlockedExtension | UploadedFileUnsafeFilename
+UploadValidationFailure = (
+    UploadedFileTooLarge | UploadedFileExtensionNotAllowed | UploadedFileBlockedExtension | UploadedFileUnsafeFilename
+)
 
 
 def _file_size_bytes(file_source: bytes | BinaryIO) -> int:
@@ -81,19 +87,35 @@ def check_upload_file_size(
     return None
 
 
+def _normalized_extension(filename: str | None) -> str:
+    if not filename:
+        return ""
+    try:
+        return Path(safe_filename(filename)).suffix.lower()
+    except ValueError:
+        return ""
+
+
+def check_allowed_extension(
+    filename: str | None,
+    allowed_extensions: tuple[str, ...] | None,
+) -> UploadedFileExtensionNotAllowed | None:
+    if allowed_extensions is None:
+        return None
+    extension: Final = _normalized_extension(filename)
+    normalized_allowed: Final = frozenset(item.lower() for item in allowed_extensions)
+    if extension and extension in normalized_allowed:
+        return None
+    return UploadedFileExtensionNotAllowed(extension=extension)
+
+
 def check_blocked_extension(
     filename: str | None,
-    blocked_extensions: tuple[str, ...],
+    blocked_extensions: tuple[str, ...] | None,
 ) -> UploadedFileBlockedExtension | None:
-    if not blocked_extensions or not filename:
+    if not blocked_extensions:
         return None
-    try:
-        extension: Final = Path(safe_filename(filename)).suffix.lower()
-    except ValueError:
-        return None
-    # The uploaded name's extension is normalized above; blocked_extensions comes
-    # straight from config.yaml or the DB and is normalized here too, so a
-    # differently-cased entry (".EXE") still catches a lowercase upload.
+    extension: Final = _normalized_extension(filename)
     normalized_blocked: Final = frozenset(item.lower() for item in blocked_extensions)
     if extension and extension in normalized_blocked:
         return UploadedFileBlockedExtension(extension=extension)
@@ -128,6 +150,17 @@ def raise_upload_validation_failure(failure: UploadValidationFailure) -> NoRetur
                 param="file",
                 code=413,
             )
+        case UploadedFileExtensionNotAllowed(extension=extension):
+            raise ProxyException(
+                message=(
+                    (f"File extension '{extension}'" if extension else "A file without an extension")
+                    + " is not in this proxy's allowed_file_extensions setting. "
+                    "The file was not forwarded to the provider."
+                ),
+                type="invalid_request_error",
+                param="file",
+                code=400,
+            )
         case UploadedFileBlockedExtension(extension=extension):
             raise ProxyException(
                 message=(

litellm/proxy/proxy_server.py

@@ -250,7 +250,7 @@ import litellm._redis
 from litellm import Router
 from litellm._logging import _redact_string, verbose_proxy_logger, verbose_router_logger
 from litellm.caching.caching import DualCache, RedisCache
-from litellm.caching.redis_cache import RedisCircuitBreakerOpenError
+from litellm.caching.redis_cache import RedisCircuitBreakerOpenError, is_redis_timeout_failure
 from litellm.caching.redis_cluster_cache import RedisClusterCache
 from litellm.constants import (
     _REALTIME_BODY_CACHE_SIZE,
@@ -365,6 +365,7 @@ from litellm.proxy.common_utils.healthy_model_filter import (
     get_hidden_unhealthy_model_names,
     is_healthy_only_listing_default,
 )
+from litellm.proxy.common_utils.html_forms.default_credentials_hint import should_hide_default_credentials_hint
 from litellm.proxy.common_utils.html_forms.ui_login import build_ui_login_form
 from litellm.proxy.common_utils.http_parsing_utils import (
     _read_request_body,
@@ -3450,8 +3451,10 @@ async def _invalidate_spend_counter(counter_key: str):
 async def _apply_spend_counter_increments(pending: Sequence[PendingSpendIncrement]) -> None:
     try:
         await increment_spend_counters_pipeline(pending=pending)
-    except RedisCircuitBreakerOpenError:
-        return
+    except Exception as e:
+        if isinstance(e, RedisCircuitBreakerOpenError) or is_redis_timeout_failure(e):
+            return
+        raise
 
 
 async def increment_spend_counters_pipeline(pending: Sequence[PendingSpendIncrement]) -> None:
@@ -7065,6 +7068,9 @@ class ProxyConfig:
         if "max_file_size_mb" not in self._yaml_general_settings_keys:
             general_settings["max_file_size_mb"] = _general_settings.get("max_file_size_mb")
 
+        if "allowed_file_extensions" not in self._yaml_general_settings_keys:
+            general_settings["allowed_file_extensions"] = _general_settings.get("allowed_file_extensions")
+
         if "blocked_file_extensions" not in self._yaml_general_settings_keys:
             general_settings["blocked_file_extensions"] = _general_settings.get("blocked_file_extensions")
 
@@ -15033,6 +15039,13 @@ def _get_proxy_model_info(model: dict) -> dict:
     return _translate_model_name_for_response(model)
 
 
+def _model_info_json_response(data: Sequence[Mapping[str, object]] | Mapping[str, object]) -> Response:
+    return Response(
+        content=orjson.dumps({"data": data}, default=jsonable_encoder, option=orjson.OPT_NON_STR_KEYS),
+        media_type="application/json",
+    )
+
+
 @router.get(
     "/model/info",
     tags=["model management"],
@@ -15080,7 +15093,7 @@ async def model_info_v1(
     `model_info.direct_access` when the proxy database is connected.
 
     Returns:
-        Returns a dictionary containing information about each model.
+        A JSON response whose `data` list holds one entry per model.
 
     Example Response:
     ```json
@@ -15128,7 +15141,7 @@ async def model_info_v1(
             deployment_dict=_deployment_info_dict,
             excluded_keys={"litellm_credential_name"},
         )
-        return {"data": _deployment_info_dict}
+        return _model_info_json_response(_deployment_info_dict)
 
     if llm_model_list is None:
         raise HTTPException(
@@ -15179,7 +15192,7 @@ async def model_info_v1(
                     llm_router=llm_router,
                     user_api_key_dict=user_api_key_dict,
                 )
-        return {"data": single_model_list}
+        return _model_info_json_response(single_model_list)
 
     # Return router deployments (same source as /v2/model/info), not wildcard-
     # expanded model names from get_complete_model_list(). Team-scoped rows
@@ -15247,7 +15260,7 @@ async def model_info_v1(
     visible_models: Final = [model for model in all_models if model.get("model_name") not in hidden_names]
 
     verbose_proxy_logger.debug("all_models: %s", visible_models)
-    return {"data": visible_models}
+    return _model_info_json_response(visible_models)
 
 
 @router.get(
@@ -15816,10 +15829,7 @@ async def fallback_login(request: Request):
 
     from fastapi.responses import HTMLResponse
 
-    hide_default_credentials_hint: Final = (
-        os.getenv("LITELLM_HIDE_DEFAULT_CREDENTIALS_HINT", "false").lower() == "true"
-        or general_settings.get("hide_default_credentials_hint", False) is True
-    )
+    hide_default_credentials_hint: Final = should_hide_default_credentials_hint(general_settings)
     return HTMLResponse(
         content=build_ui_login_form(
             show_deprecation_banner=False,
@@ -17036,6 +17046,7 @@ _GENERAL_SETTINGS_CONFIG_LIST_FIELD_TYPES: Final[Mapping[str, str]] = MappingPro
         "max_request_size_mb": "Integer",
         "max_batch_file_size_mb": "Integer",
         "max_file_size_mb": "Integer",
+        "allowed_file_extensions": "List",
         "blocked_file_extensions": "List",
         "max_response_size_mb": "Integer",
         "proxy_config_reload_interval_seconds": "Integer",

litellm/responses/litellm_completion_transformation/transformation.py

@@ -43,6 +43,7 @@ from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging
 from litellm.responses.litellm_completion_transformation.session_handler import (
     ResponsesSessionHandler,
 )
+from litellm.types.llms.base import CachedTokensDetails
 from litellm.types.llms.openai import (
     AllMessageValues,
     ChatCompletionAssistantMessage,
@@ -2816,27 +2817,24 @@ class LiteLLMCompletionResponsesConfig:
         # Translate prompt_tokens_details to input_tokens_details
         if hasattr(usage, "prompt_tokens_details") and usage.prompt_tokens_details is not None:
             prompt_details: Final = usage.prompt_tokens_details
-            input_details_dict: Final[dict[str, int]] = {}
-
-            if hasattr(prompt_details, "cached_tokens") and prompt_details.cached_tokens is not None:
-                input_details_dict["cached_tokens"] = prompt_details.cached_tokens
-            else:
-                input_details_dict["cached_tokens"] = 0
-
-            if hasattr(prompt_details, "text_tokens") and prompt_details.text_tokens is not None:
-                input_details_dict["text_tokens"] = prompt_details.text_tokens
-
-            if hasattr(prompt_details, "audio_tokens") and prompt_details.audio_tokens is not None:
-                input_details_dict["audio_tokens"] = prompt_details.audio_tokens
-
-            cache_write_tokens = getattr(prompt_details, "cache_write_tokens", None) or getattr(
+            cached_tokens_details: Final = getattr(prompt_details, "cached_tokens_details", None)
+            cache_write_tokens: Final = getattr(prompt_details, "cache_write_tokens", None) or getattr(
                 prompt_details, "cache_creation_tokens", None
             )
-            if cache_write_tokens is not None:
-                input_details_dict["cache_write_tokens"] = cache_write_tokens
-
-            if input_details_dict:
-                response_usage.input_tokens_details = InputTokensDetails(**input_details_dict)
+            cache_write_extra: Final[Mapping[str, int]] = (
+                MappingProxyType({"cache_write_tokens": cache_write_tokens})
+                if cache_write_tokens is not None
+                else MappingProxyType({})
+            )
+            response_usage.input_tokens_details = InputTokensDetails(
+                cached_tokens=prompt_details.cached_tokens if prompt_details.cached_tokens is not None else 0,
+                text_tokens=prompt_details.text_tokens,
+                audio_tokens=prompt_details.audio_tokens,
+                cached_tokens_details=(
+                    cached_tokens_details if isinstance(cached_tokens_details, CachedTokensDetails) else None
+                ),
+                **cache_write_extra,
+            )
 
         # Translate completion_tokens_details to output_tokens_details
         if hasattr(usage, "completion_tokens_details") and usage.completion_tokens_details is not None:

litellm/responses/utils.py

@@ -1179,6 +1179,9 @@ class ResponseAPILoggingUtils:
                     audio_tokens=getattr(response_api_usage.input_tokens_details, "audio_tokens", None),
                     text_tokens=getattr(response_api_usage.input_tokens_details, "text_tokens", None),
                     image_tokens=getattr(response_api_usage.input_tokens_details, "image_tokens", None),
+                    cached_tokens_details=getattr(
+                        response_api_usage.input_tokens_details, "cached_tokens_details", None
+                    ),
                     cache_write_tokens=getattr(response_api_usage.input_tokens_details, "cache_write_tokens", None),
                 )
         completion_tokens_details: CompletionTokensDetailsWrapper | None = None

litellm/router.py

@@ -96,7 +96,6 @@ from litellm.litellm_core_utils.request_timeout_resolver import (
 from litellm.litellm_core_utils.secret_redaction import redact_string
 from litellm.litellm_core_utils.sensitive_data_masker import (
     SensitiveDataMasker,
-    mask_credentials_in_payload,
     mask_sensitive_structure,
 )
 from litellm.litellm_core_utils.token_counter import offload_token_count
@@ -623,20 +622,6 @@ def _replay_live_router_model_cost() -> None:
 set_live_deployment_replay(_replay_live_router_model_cost)
 
 
-# Kwargs that carry no signal about the failed attempt, so log_retry drops them from a
-# breadcrumb entirely: the request payload, the proxy's snapshot of the inbound request (its body
-# aliases the live request metadata, earlier breadcrumbs included, so copying it would nest every
-# breadcrumb inside the next one), and the router-internal walk state. Credentials are handled
-# separately by mask_credentials_in_payload, which scrubs credential-named values from whatever
-# kwargs remain rather than trying to enumerate every credential-bearing key here.
-RETRY_BREADCRUMB_EXCLUDED_KWARGS: Final = frozenset(
-    (
-        "messages",
-        "original_function",
-        "attempted_targets",
-        "proxy_server_request",
-    )
-)
 RETRY_BREADCRUMB_LIMIT: Final = 4
 
 
@@ -8374,31 +8359,30 @@ class Router:
 
     def log_retry(self, kwargs: dict, e: Exception) -> dict:
         """
-        When a retry or fallback happens, log the details of the just failed model call - similar to Sentry breadcrumbing
+        When a retry or fallback happens, record which model group, deployment and attempt just failed and why
         """
+        from litellm.types.router import RetryAttemptRecord
+
         _metadata_var: Final = "litellm_metadata" if "litellm_metadata" in kwargs else "metadata"
         request_metadata: Final[Mapping[str, object]] = kwargs[_metadata_var]
-        attempt_kwargs: Final = MappingProxyType(
-            {k: v for k, v in kwargs.items() if k != _metadata_var and k not in RETRY_BREADCRUMB_EXCLUDED_KWARGS}
-        )
-        attempt_metadata: Final = MappingProxyType(
-            {k: v for k, v in request_metadata.items() if k != "previous_models"}
-        )
-        previous_model: Final = MappingProxyType(
-            {
-                "exception_type": type(e).__name__,
-                "exception_string": str(e),
-                **attempt_kwargs,
-                _metadata_var: attempt_metadata,
-            }
-        )
+        model_group: Final = kwargs.get("model")
+        model_info: Final = request_metadata.get("model_info")
+        deployment_id: Final = model_info.get("id") if isinstance(model_info, Mapping) else None
+        attempted_retries: Final = request_metadata.get("attempted_retries")
+        attempt_record: Final[RetryAttemptRecord] = {
+            "model_group": model_group if isinstance(model_group, str) else None,
+            "deployment_id": deployment_id if isinstance(deployment_id, str) else None,
+            "exception_type": type(e).__name__,
+            "exception_string": str(e),
+            "attempted_retries": attempted_retries if type(attempted_retries) is int else None,
+        }
         earlier_breadcrumbs: Final = request_metadata.get("previous_models")
         kept_breadcrumbs: Final[tuple[object, ...]] = (
             tuple(earlier_breadcrumbs)[-(RETRY_BREADCRUMB_LIMIT - 1) :]
             if isinstance(earlier_breadcrumbs, (list, tuple))
             else ()
         )
-        breadcrumbs: Final = (*kept_breadcrumbs, mask_credentials_in_payload(previous_model))
+        breadcrumbs: Final = (*kept_breadcrumbs, attempt_record)
         kwargs[_metadata_var]["previous_models"] = breadcrumbs  # rebind-ok: the logging object already holds this dict
         return kwargs
 
@@ -13878,6 +13862,7 @@ class Router:
                 cooldown_time=_cooldown_time,
                 enable_pre_call_checks=self.enable_pre_call_checks,
                 cooldown_list=_cooldown_list,
+                model_ids=model_ids,
             )
 
         if strategy == "simple-shuffle":
@@ -13910,6 +13895,7 @@ class Router:
                 cooldown_time=_cooldown_time,
                 enable_pre_call_checks=self.enable_pre_call_checks,
                 cooldown_list=_cooldown_list,
+                model_ids=model_ids,
             )
         self._override_selector_pre_call_check(strategy, strategy_selector, deployment)
         verbose_router_logger.info(
@@ -13987,6 +13973,11 @@ class Router:
                 model=model,
                 llm_provider="",
             )
+        pass_through_model_ids: Final = tuple(
+            deployment["model_info"]["id"]
+            for deployment in pass_through_deployments
+            if "id" in deployment.get("model_info", {})
+        )
 
         # 4. Apply health-check and cooldown filtering
         parent_otel_span: Final[Span | None] = _get_parent_otel_span_from_kwargs(request_kwargs)
@@ -14024,6 +14015,7 @@ class Router:
                 cooldown_time=_cooldown_time,
                 enable_pre_call_checks=self.enable_pre_call_checks,
                 cooldown_list=_cooldown_list,
+                model_ids=pass_through_model_ids,
             )
 
         # 6. Apply load balancing strategy
@@ -14057,6 +14049,7 @@ class Router:
                 cooldown_time=_cooldown_time,
                 enable_pre_call_checks=self.enable_pre_call_checks,
                 cooldown_list=_cooldown_list,
+                model_ids=model_ids,
             )
         self._override_selector_pre_call_check(strategy, strategy_selector, deployment)
 

litellm/router_utils/handle_error.py

@@ -93,4 +93,5 @@ async def async_raise_no_deployment_exception(
         cooldown_time=_cooldown_time,
         enable_pre_call_checks=litellm_router_instance.enable_pre_call_checks,
         cooldown_list=cooldown_list_ids,
+        model_ids=model_ids,
     )

litellm/rust_bridge/lifecycle.py

@@ -99,23 +99,13 @@ def setup(
 
 def check_limits(kwargs: Mapping[str, object]) -> None:
     import litellm
+    from litellm.litellm_core_utils.core_helpers import max_retries_per_request_hit
 
     current_cost: Final = litellm._current_cost  # pyright: ignore[reportPrivateUsage]  # shared SDK budget counter has no public accessor
     if litellm.max_budget and current_cost > litellm.max_budget:
         raise litellm.BudgetExceededError(current_cost=current_cost, max_budget=litellm.max_budget)
-    metadata: Final = kwargs.get("metadata")
-    if isinstance(metadata, Mapping):
-        typed_metadata: Final = cast(  # cast-ok: runtime Mapping check establishes read-only metadata
-            Mapping[str, object], metadata
-        )
-        previous: Final = typed_metadata.get("previous_models")
-        if (
-            isinstance(previous, list)
-            and litellm.num_retries_per_request is not None
-            and len(cast(list[object], previous))  # cast-ok: runtime list check establishes the retry history
-            >= litellm.num_retries_per_request
-        ):
-            raise RuntimeError("Max retries per request hit!")
+    if max_retries_per_request_hit(kwargs, litellm.num_retries_per_request):
+        raise RuntimeError("Max retries per request hit!")
 
 
 def finalize(

litellm/types/llms/base.py

@@ -75,3 +75,9 @@ class HiddenParams(OpenAIObject):
         data: Final = super().model_dump(**kwargs)
         data["_response_ms"] = self._response_ms
         return data
+
+
+class CachedTokensDetails(BaseModel):
+    text_tokens: int | None = None
+    audio_tokens: int | None = None
+    image_tokens: int | None = None

litellm/types/llms/openai.py

@@ -91,6 +91,8 @@ from litellm.types.responses.main import (
     OutputImageGenerationCall,
 )
 
+from .base import CachedTokensDetails
+
 FileContent = IO[bytes] | bytes | PathLike
 
 FileTypes = (
@@ -1288,6 +1290,7 @@ class OutputTokensDetails(BaseLiteLLMOpenAIResponseObject):
 class InputTokensDetails(BaseLiteLLMOpenAIResponseObject):
     audio_tokens: int | None = None
     cached_tokens: int = 0
+    cached_tokens_details: CachedTokensDetails | None = None
     text_tokens: int | None = None
 
     model_config = {"extra": "allow"}
@@ -2254,10 +2257,17 @@ class OpenAIRealtimeInputAudioTranscriptionCompleted(TypedDict):
     usage: NotRequired[ReadOnly[Mapping[str, object]]]
 
 
+class OpenAIRealtimeCachedTokensDetails(TypedDict, total=False):
+    text_tokens: ReadOnly[int]
+    audio_tokens: ReadOnly[int]
+    image_tokens: ReadOnly[int]
+
+
 class OpenAIRealtimeUsageTokenDetails(TypedDict):
     audio_tokens: ReadOnly[int]
     text_tokens: ReadOnly[int]
     cached_tokens: NotRequired[ReadOnly[int]]
+    cached_tokens_details: NotRequired[ReadOnly[OpenAIRealtimeCachedTokensDetails]]
 
 
 class OpenAIRealtimeResponseUsage(TypedDict):

litellm/types/proxy/auth/auth_checks.py

@@ -0,0 +1,8 @@
+"""Failure values raised by `litellm/proxy/auth/auth_checks.py`. Kept free of `litellm` imports so any proxy module can import them without joining the `litellm.proxy` import cycle."""
+
+
+class UserNotFoundError(ValueError):
+    """The user row is provably absent, as opposed to merely unreadable, so a caller that reads a missing row as no user-level limits can key on it without also swallowing a database that would not answer."""
+
+    def __init__(self, user_id: str) -> None:
+        super().__init__(f"User doesn't exist in db. 'user_id'={user_id}. Create user via `/user/new` call.")

litellm/types/router.py

@@ -645,6 +645,7 @@ class RouterErrors(enum.Enum):
 
     user_defined_ratelimit_error = "Deployment over user-defined ratelimit."
     no_deployments_available = "No deployments available for selected model"
+    all_deployments_in_cooldown = "All deployments for selected model are in cooldown"
     no_deployments_with_tag_routing = "Not allowed to access model due to tags configuration"
     no_deployments_with_provider_budget_routing = "No deployments available - crossed budget"
     no_healthy_deployments = "There are no healthy deployments for this model"
@@ -868,6 +869,11 @@ class RouterRateLimitErrorBasic(ValueError):
         super().__init__(_message)
 
 
+class RouterErrorTypes(str, enum.Enum):
+    rate_limit_error = "rate_limit_error"
+    all_deployments_in_cooldown = "all_deployments_in_cooldown"
+
+
 class RouterRateLimitError(ValueError):
     def __init__(
         self,
@@ -875,12 +881,25 @@ class RouterRateLimitError(ValueError):
         cooldown_time: float,
         enable_pre_call_checks: bool,
         cooldown_list: list,
+        model_ids: Sequence[str] = (),
     ) -> None:
         self.model = model
         self.cooldown_time = cooldown_time
         self.enable_pre_call_checks = enable_pre_call_checks
         self.cooldown_list = cooldown_list
-        _message = f"{RouterErrors.no_deployments_available.value}, Try again in {cooldown_time} seconds. Passed model={model}. pre-call-checks={enable_pre_call_checks}, cooldown_list={cooldown_list}"
+        self.all_deployments_in_cooldown = bool(model_ids) and frozenset(model_ids) <= frozenset(cooldown_list)
+        self.type = (
+            RouterErrorTypes.all_deployments_in_cooldown.value
+            if self.all_deployments_in_cooldown
+            else RouterErrorTypes.rate_limit_error.value
+        )
+        _reason: Final = (
+            f" {RouterErrors.all_deployments_in_cooldown.value}." if self.all_deployments_in_cooldown else ""
+        )
+        _message: Final = (
+            f"{RouterErrors.no_deployments_available.value}, Try again in {cooldown_time} seconds.{_reason} "
+            f"Passed model={model}. pre-call-checks={enable_pre_call_checks}, cooldown_list={cooldown_list}"
+        )
         super().__init__(_message)
 
 
@@ -889,6 +908,14 @@ class RouterModelGroupAliasItem(TypedDict):
     hidden: bool  # if 'True', don't return on `.get_model_list`
 
 
+class RetryAttemptRecord(TypedDict):
+    model_group: ReadOnly[str | None]
+    deployment_id: ReadOnly[str | None]
+    exception_type: ReadOnly[str]
+    exception_string: ReadOnly[str]
+    attempted_retries: ReadOnly[int | None]
+
+
 VALID_LITELLM_ENVIRONMENTS = [
     "development",
     "staging",

litellm/types/utils.py

@@ -48,6 +48,7 @@ from litellm._logging import verbose_logger
 from litellm._uuid import uuid
 from litellm.types.llms.base import (
     BaseLiteLLMOpenAIResponseObject,
+    CachedTokensDetails,
     LiteLLMPydanticObjectBase,
 )
 from litellm.types.mcp import MCPServerCostInfo
@@ -252,6 +253,7 @@ class ModelInfoBase(ProviderSpecificModelInfo, total=False):
     cache_creation_input_token_cost_priority: float | None  # OpenAI priority service tier pricing
     cache_creation_input_token_cost_ultrafast: ReadOnly[float | None]  # OpenAI ultrafast service tier pricing
     cache_read_input_token_cost: float | None
+    cache_read_input_audio_token_cost: ReadOnly[float | None]
     cache_read_input_token_cost_flex: float | None  # OpenAI flex service tier pricing
     cache_read_input_token_cost_priority: float | None  # OpenAI priority service tier pricing
     cache_read_input_token_cost_ultrafast: ReadOnly[float | None]  # OpenAI ultrafast service tier pricing
@@ -1710,6 +1712,9 @@ class PromptTokensDetailsWrapper(
     cache_creation_token_details: CacheCreationTokenDetails | None = None
     """Details of cache creation tokens sent to the model. Used for tracking 5m/1h cache creation tokens for Anthropic prompt caching."""
 
+    cached_tokens_details: CachedTokensDetails | None = None
+    """Details of cached (cache-hit) tokens sent to the model. OpenAI realtime naming; carries the per-modality cache-read split."""
+
     def __setattr__(self, name: str, value: object) -> None:
         super().__setattr__(name, value)
         if name == "cache_write_tokens":
@@ -1756,6 +1761,8 @@ class PromptTokensDetailsWrapper(
             del self.cache_creation_tokens
         if self.cache_creation_token_details is None:
             del self.cache_creation_token_details
+        if self.cached_tokens_details is None:
+            del self.cached_tokens_details
 
 
 class ServerToolUse(BaseModel):

litellm/utils.py

@@ -84,6 +84,7 @@ from litellm.constants import (
 from litellm.litellm_core_utils.core_helpers import normalize_drop_params
 from litellm.litellm_core_utils.fallback_generalizations import (
     match_capability_generalizations,
+    match_fill_missing_generalizations,
 )
 from litellm.litellm_core_utils.sensitive_data_masker import redact_credentials_in_payload
 
@@ -254,6 +255,7 @@ from litellm.types.utils import (
 )
 
 _CALL_TYPE_ENUM_MAP: Final[dict] = {ct.value: ct for ct in CallTypes}
+_BACKFILL_MODES: Final = frozenset({"chat", "responses"})
 
 # +-----------------------------------------------+
 # |                                               |
@@ -1260,15 +1262,6 @@ async def _client_async_logging_helper(
             async_coroutine=logging_obj.async_success_handler(result=result, start_time=start_time, end_time=end_time)
         )
 
-        ################################################
-        # Sync Logging Worker
-        ################################################
-        logging_obj.handle_sync_success_callbacks_for_async_calls(
-            result=result,
-            start_time=start_time,
-            end_time=end_time,
-        )
-
 
 def _get_wrapper_num_retries(kwargs: dict[str, Any], exception: Exception) -> tuple[int | None, dict[str, Any]]:
     """
@@ -1500,6 +1493,8 @@ def post_call_processing(
 
 
 def client(original_function):
+    from litellm.litellm_core_utils.core_helpers import max_retries_per_request_hit
+
     Rules: Final = litellm_utils.Rules
     rules_obj: Final = Rules()
 
@@ -1510,12 +1505,8 @@ def client(original_function):
         call_type = original_function.__name__
         if _is_async_request(kwargs):
             # [OPTIONAL] CHECK MAX RETRIES / REQUEST
-            if litellm.num_retries_per_request is not None:
-                # check if previous_models passed in as ['litellm_params']['metadata]['previous_models']
-                previous_models = (kwargs.get("metadata") or {}).get("previous_models", None)
-                if previous_models is not None:
-                    if litellm.num_retries_per_request <= len(previous_models):
-                        raise Exception("Max retries per request hit!")
+            if max_retries_per_request_hit(kwargs, litellm.num_retries_per_request):
+                raise Exception("Max retries per request hit!")
 
             # MODEL CALL
             result = original_function(*args, **kwargs)
@@ -1574,12 +1565,8 @@ def client(original_function):
                     )
 
             # [OPTIONAL] CHECK MAX RETRIES / REQUEST
-            if litellm.num_retries_per_request is not None:
-                # check if previous_models passed in as ['litellm_params']['metadata]['previous_models']
-                previous_models = (kwargs.get("metadata") or {}).get("previous_models", None)
-                if previous_models is not None:
-                    if litellm.num_retries_per_request <= len(previous_models):
-                        raise Exception("Max retries per request hit!")
+            if max_retries_per_request_hit(kwargs, litellm.num_retries_per_request):
+                raise Exception("Max retries per request hit!")
 
             # [OPTIONAL] CHECK CACHE
             print_verbose(
@@ -5819,6 +5806,14 @@ def _get_model_info_helper(
                     ):
                         _model_info = None
 
+            if _model_info is not None and key is not None and _model_info.get("mode", "chat") in _BACKFILL_MODES:
+                fill_missing: Final = match_fill_missing_generalizations(key, _model_info.get("litellm_provider", ""))
+                if fill_missing is not None:
+                    _model_info = {
+                        **{k: v for k, v in fill_missing.items() if k not in _model_info},
+                        **_model_info,
+                    }
+
             if _model_info is None:
                 generalization: Final = _get_model_info_from_generalization(
                     model=model,
@@ -5882,6 +5877,7 @@ def _get_model_info_helper(
                     "cache_creation_input_token_cost_ultrafast", None
                 ),
                 cache_read_input_token_cost=_model_info.get("cache_read_input_token_cost", None),
+                cache_read_input_audio_token_cost=_model_info.get("cache_read_input_audio_token_cost", None),
                 prompt_cache_min_tokens=_model_info.get("prompt_cache_min_tokens", None),
                 cache_read_input_token_cost_above_200k_tokens=_model_info.get(
                     "cache_read_input_token_cost_above_200k_tokens", None

model_prices_and_context_window.json

@@ -5513,7 +5513,8 @@
     },
     "azure/gpt-realtime-2025-08-28": {
         "cache_creation_input_audio_token_cost": 4e-06,
-        "cache_read_input_token_cost": 4e-06,
+        "cache_read_input_audio_token_cost": 4e-07,
+        "cache_read_input_token_cost": 4e-07,
         "deprecation_date": "2027-03-02",
         "input_cost_per_audio_token": 3.2e-05,
         "input_cost_per_image_token": 5e-06,
@@ -5546,7 +5547,8 @@
     },
     "azure/gpt-realtime-1.5-2026-02-23": {
         "cache_creation_input_audio_token_cost": 4e-06,
-        "cache_read_input_token_cost": 4e-06,
+        "cache_read_input_audio_token_cost": 4e-07,
+        "cache_read_input_token_cost": 4e-07,
         "deprecation_date": "2027-08-24",
         "input_cost_per_audio_token": 3.2e-05,
         "input_cost_per_image_token": 5e-06,
@@ -5683,6 +5685,7 @@
     },
     "azure/gpt-realtime-mini": {
         "cache_creation_input_audio_token_cost": 3e-07,
+        "cache_read_input_audio_token_cost": 3e-07,
         "cache_read_input_token_cost": 6e-08,
         "input_cost_per_audio_token": 1e-05,
         "input_cost_per_image_token": 8e-07,
@@ -5715,6 +5718,7 @@
     },
     "azure/gpt-realtime-mini-2025-10-06": {
         "cache_creation_input_audio_token_cost": 3e-07,
+        "cache_read_input_audio_token_cost": 3e-07,
         "cache_read_input_token_cost": 6e-08,
         "input_cost_per_audio_token": 1e-05,
         "input_cost_per_image_token": 8e-07,
@@ -7409,6 +7413,80 @@
         "supports_web_search": true,
         "supports_xhigh_reasoning_effort": true
     },
+    "azure/gpt-chat-latest": {
+        "cache_read_input_token_cost": 5e-07,
+        "deprecation_date": "2026-12-02",
+        "input_cost_per_token": 5e-06,
+        "litellm_provider": "azure",
+        "max_input_tokens": 272000,
+        "max_output_tokens": 128000,
+        "max_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 3e-05,
+        "reasoning_effort_levels": [
+            "medium"
+        ],
+        "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/",
+        "supported_endpoints": [
+            "/v1/chat/completions",
+            "/v1/responses"
+        ],
+        "supported_modalities": [
+            "text",
+            "image"
+        ],
+        "supported_output_modalities": [
+            "text"
+        ],
+        "supports_function_calling": true,
+        "supports_native_streaming": true,
+        "supports_parallel_function_calling": true,
+        "supports_pdf_input": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "supports_vision": true,
+        "supports_web_search": true
+    },
+    "azure/chat-latest": {
+        "cache_read_input_token_cost": 5e-07,
+        "deprecation_date": "2026-12-02",
+        "input_cost_per_token": 5e-06,
+        "litellm_provider": "azure",
+        "max_input_tokens": 272000,
+        "max_output_tokens": 128000,
+        "max_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 3e-05,
+        "reasoning_effort_levels": [
+            "medium"
+        ],
+        "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/",
+        "supported_endpoints": [
+            "/v1/chat/completions",
+            "/v1/responses"
+        ],
+        "supported_modalities": [
+            "text",
+            "image"
+        ],
+        "supported_output_modalities": [
+            "text"
+        ],
+        "supports_function_calling": true,
+        "supports_native_streaming": true,
+        "supports_parallel_function_calling": true,
+        "supports_pdf_input": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "supports_vision": true,
+        "supports_web_search": true
+    },
     "azure/us/gpt-5.6": {
         "cache_creation_input_token_cost": 6.875e-06,
         "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05,
@@ -7675,6 +7753,43 @@
         "supports_web_search": true,
         "supports_xhigh_reasoning_effort": true
     },
+    "azure/us/gpt-chat-latest": {
+        "cache_read_input_token_cost": 5.5e-07,
+        "deprecation_date": "2026-12-02",
+        "input_cost_per_token": 5.5e-06,
+        "litellm_provider": "azure",
+        "max_input_tokens": 272000,
+        "max_output_tokens": 128000,
+        "max_tokens": 128000,
+        "mode": "chat",
+        "output_cost_per_token": 3.3e-05,
+        "reasoning_effort_levels": [
+            "medium"
+        ],
+        "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/",
+        "supported_endpoints": [
+            "/v1/chat/completions",
+            "/v1/responses"
+        ],
+        "supported_modalities": [
+            "text",
+            "image"
+        ],
+        "supported_output_modalities": [
+            "text"
+        ],
+        "supports_function_calling": true,
+        "supports_native_streaming": true,
+        "supports_parallel_function_calling": true,
+        "supports_pdf_input": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "supports_vision": true,
+        "supports_web_search": true
+    },
     "azure/eu/gpt-5.6": {
         "cache_creation_input_token_cost": 6.875e-06,
         "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05,
@@ -23113,34 +23228,6 @@
         "output_cost_per_token": 0.0,
         "source": "https://fireworks.ai/pricing"
     },
-    "friendliai/meta-llama-3.1-70b-instruct": {
-        "input_cost_per_token": 6e-07,
-        "litellm_provider": "friendliai",
-        "max_input_tokens": 8192,
-        "max_output_tokens": 8192,
-        "max_tokens": 8192,
-        "mode": "chat",
-        "output_cost_per_token": 6e-07,
-        "supports_function_calling": true,
-        "supports_parallel_function_calling": true,
-        "supports_response_schema": true,
-        "supports_system_messages": true,
-        "supports_tool_choice": true
-    },
-    "friendliai/meta-llama-3.1-8b-instruct": {
-        "input_cost_per_token": 1e-07,
-        "litellm_provider": "friendliai",
-        "max_input_tokens": 8192,
-        "max_output_tokens": 8192,
-        "max_tokens": 8192,
-        "mode": "chat",
-        "output_cost_per_token": 1e-07,
-        "supports_function_calling": true,
-        "supports_parallel_function_calling": true,
-        "supports_response_schema": true,
-        "supports_system_messages": true,
-        "supports_tool_choice": true
-    },
     "friendliai/zai-org/GLM-5.3-Flash": {
         "litellm_provider": "friendliai",
         "max_input_tokens": 1048576,
@@ -32938,6 +33025,7 @@
     },
     "gpt-realtime": {
         "cache_creation_input_audio_token_cost": 4e-07,
+        "cache_read_input_audio_token_cost": 4e-07,
         "cache_read_input_token_cost": 4e-07,
         "deprecation_date": "2027-01-20",
         "input_cost_per_audio_token": 3.2e-05,
@@ -32972,6 +33060,7 @@
     },
     "gpt-realtime-1.5": {
         "cache_creation_input_audio_token_cost": 4e-07,
+        "cache_read_input_audio_token_cost": 4e-07,
         "cache_read_input_token_cost": 4e-07,
         "input_cost_per_audio_token": 3.2e-05,
         "input_cost_per_image_token": 5e-06,
@@ -33146,6 +33235,7 @@
     },
     "gpt-realtime-2025-08-28": {
         "cache_creation_input_audio_token_cost": 4e-07,
+        "cache_read_input_audio_token_cost": 4e-07,
         "cache_read_input_token_cost": 4e-07,
         "deprecation_date": "2027-01-20",
         "input_cost_per_audio_token": 3.2e-05,
@@ -57947,8 +58037,9 @@
             },
             {
                 "name": "claude-adaptive-thinking",
-                "pattern": "claude-[a-z]+-(?:4[-._](?:[6-9]|[1-9]\\d)(?!\\d)|(?:[5-9]|[1-9]\\d)(?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)",
-                "description": "Claude at version 4.6 or higher, in any id shape that contains claude-<family>-: minors 4.6 through 4.99, any later major-minor, and bare 5+ majors so a new family shaped like claude-fable-5 matches. Requiring the claude- prefix keeps non-Claude names such as team-sonnet-5-1 out. The minor is capped at two digits so an 8-digit date suffix such as claude-opus-4-20250514 is never read as a >= 4.6 minor. Turns on adaptive thinking for new versions and new families with no code change.",
+                "pattern": "claude-[a-z]+-(?:4[-._](?:[6-9]|[1-9]\\d)(?!\\d)|[5-9](?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)",
+                "fill_missing_for_providers": ["anthropic", "azure_ai", "bedrock", "bedrock_converse", "vertex_ai-anthropic_models"],
+                "description": "Claude at version 4.6 or higher, in any id shape that contains claude-<family>-: minors 4.6 through 4.99, any later major-minor, and bare 5+ majors so a new family shaped like claude-fable-5 matches. Requiring the claude- prefix keeps non-Claude names such as team-sonnet-5-1 out. The minor is capped at two digits so an 8-digit date suffix such as claude-opus-4-20250514 is never read as a >= 4.6 minor. Two-digit majors are deliberately not matched so ids like claude-opus-41 (4.1) are not read as major 41. Turns on adaptive thinking for new versions and new families with no code change.",
                 "model_info": {
                     "supports_adaptive_thinking": true
                 }
@@ -57956,6 +58047,7 @@
             {
                 "name": "claude-legacy-thinking",
                 "pattern": "claude-[a-z]+-4[-._]6(?!\\d)",
+                "fill_missing_for_providers": ["anthropic", "azure_ai", "bedrock", "bedrock_converse", "vertex_ai-anthropic_models"],
                 "description": "Claude at version 4.6 exactly, in any id shape that contains claude-<family>-4-6 (dotted and underscored minors included, dated releases such as claude-sonnet-4-6-20260219 too). The 4.6 family is adaptive-thinking yet still accepts legacy thinking.type=enabled with budget_tokens, so the caller's hard budget cap is forwarded verbatim instead of being rewritten to an uncapped output_config.effort. The lookahead keeps two-digit minors such as 4-60 from matching. 4.7+ and 5+ majors reject the legacy shape and stay on the adaptive translation.",
                 "model_info": {
                     "supports_legacy_thinking": true
@@ -57971,8 +58063,9 @@
             },
             {
                 "name": "claude-mid-conversation-system",
-                "pattern": "claude-[a-z]+-(?:4[-._](?:[89]|[1-9]\\d)(?!\\d)|(?:[5-9]|[1-9]\\d)(?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)",
-                "description": "Claude at version 4.8 or higher, in any id shape that contains claude-<family>-: minors 4.8 through 4.99, any later major-minor, and bare 5+ majors so a new family like claude-fable-5 matches. Anthropic introduced mid-conversation system messages with Opus 4.8 and every newer Claude keeps them; 4.7 and below reject the system role inside messages.",
+                "pattern": "claude-[a-z]+-(?:4[-._](?:[89]|[1-9]\\d)(?!\\d)|[5-9](?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)",
+                "fill_missing_for_providers": ["anthropic", "azure_ai", "bedrock", "bedrock_converse", "vertex_ai-anthropic_models"],
+                "description": "Claude at version 4.8 or higher, in any id shape that contains claude-<family>-: minors 4.8 through 4.99, any later major-minor, and bare 5+ majors so a new family like claude-fable-5 matches. Two-digit majors are deliberately not matched so ids like claude-opus-41 (4.1) are not read as major 41. Anthropic introduced mid-conversation system messages with Opus 4.8 and every newer Claude keeps them; 4.7 and below reject the system role inside messages.",
                 "model_info": {
                     "supports_mid_conversation_system": true
                 }
@@ -57988,6 +58081,7 @@
             {
                 "name": "openai-reasoning-family-baseline",
                 "pattern": "^(?!.*search-api)(?:[a-z0-9_.-]+/)*(?:ft:)?(?:o[1-9]\\d*(?![a-z0-9])|gpt-[5-9](?:\\.\\d+)?(?![0-9.])|(?:gpt-\\d+(?:\\.\\d+)?(?:-[a-z0-9]+)*-)?(?:codex|deep-research|chat-latest)(?![a-z0-9]))",
+                "fill_missing_for_providers": ["azure", "azure_ai", "openai"],
                 "description": "OpenAI reasoning families by id shape, under any provider namespace and with an optional ft: prefix: the o-series (o1, o3-pro, o4-mini), gpt-5 through gpt-9 majors including dotted minors and suffixed variants (gpt-5.5-cyber, gpt-6-astra), and the codex, deep-research and chat-latest lines when standalone or on a gpt base. gpt-5-search-api is excluded because it is a search-only surface. Every model here is a reasoning model, and the Responses API drops the caller's reasoning param for any mapped OpenAI model whose info lacks supports_reasoning, so an id the registry has not named yet keeps its reasoning settings instead of silently losing them. Rules lose to exact entries. Carries no mode and no pricing, so cost stays on the standard unpriced behavior.",
                 "model_info": {
                     "supports_reasoning": true

osv-scanner.toml

@@ -5,5 +5,5 @@ reason = "diskcache has no fixed release published; remove this entry once one e
 
 [[IgnoredVulns]]
 id = "GHSA-h7x2-h6g9-p789"
-ignoreUntil = 2026-09-14
-reason = "mlflow has no fixed release published; remove this entry once one exists"
+ignoreUntil = 2026-10-14
+reason = "mlflow has no fixed release published (3.16.0, 2026-09-04, and master still store gateway secret api_base unvalidated); remove this entry once one exists"

tests/proxy_behavior/auth/test_auth_object_prefetch.py

@@ -22,6 +22,11 @@ from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache
 pytestmark = pytest.mark.asyncio(loop_scope="session")
 
 
+def _frozen_cache() -> UserApiKeyCache:
+    """The org entries carry a 5s TTL; a frozen clock keeps a slow first call from expiring them mid-test."""
+    return UserApiKeyCache(in_memory_cache=InMemoryCache(clock=lambda: 1_000_000.0), redis_cache=None)
+
+
 def _dead_db() -> MagicMock:
     prisma = MagicMock(name="prisma_client")
     prisma.db.query_first = AsyncMock(return_value=None)
@@ -58,7 +63,7 @@ async def test_join_binds_the_membership_to_the_requested_team(prisma):
             data={"user_id": user_id, "team_id": team_b, "litellm_budget_table": {"connect": {"budget_id": f"b-{run}"}}}
         )
 
-        cache = UserApiKeyCache(in_memory_cache=InMemoryCache(), redis_cache=None)
+        cache = _frozen_cache()
         refs = AuthObjectRefs(user_id=user_id, team_id=team_a, membership_user_id=user_id, organization_id=org_id)
         await prefetch_auth_objects(refs=refs, user_api_key_cache=cache, prisma_client=prisma)
 
@@ -100,7 +105,7 @@ async def test_join_reads_team_model_aliases_from_the_mapped_column(prisma):
             where={"team_id": team_id}, include={"litellm_model_table": True}
         )
 
-        cache = UserApiKeyCache(in_memory_cache=InMemoryCache(), redis_cache=None)
+        cache = _frozen_cache()
         refs = AuthObjectRefs(user_id=None, team_id=team_id, membership_user_id=None, organization_id=None)
         await prefetch_auth_objects(refs=refs, user_api_key_cache=cache, prisma_client=prisma)
 
@@ -144,7 +149,7 @@ async def test_join_reads_null_nested_lists_the_way_prisma_does(prisma):
             where={"user_id_team_id": {"user_id": user_id, "team_id": team_id}}, include={"litellm_budget_table": True}
         )
 
-        cache = UserApiKeyCache(in_memory_cache=InMemoryCache(), redis_cache=None)
+        cache = _frozen_cache()
         refs = AuthObjectRefs(user_id=user_id, team_id=team_id, membership_user_id=user_id, organization_id=None)
         await prefetch_auth_objects(refs=refs, user_api_key_cache=cache, prisma_client=prisma)
 

tests/proxy_unit_tests/test_proxy_server.py

@@ -2139,7 +2139,7 @@ async def test_model_info_alias_without_prisma(hidden):
         user_api_key_dict=UserAPIKeyAuth(models=[]),
     )
 
-    models = resp["data"]
+    models = json.loads(resp.body)["data"]
 
     alias_found = any(
         m["model_name"] == model_alias
@@ -2203,7 +2203,7 @@ async def test_proxy_model_group_alias_checks(prisma_client, hidden):  # noqa: F
     resp = await model_info_v1(
         user_api_key_dict=UserAPIKeyAuth(models=[]),
     )
-    models = resp["data"]
+    models = json.loads(resp.body)["data"]
     is_model_alias_in_list = False
     for item in models:
         if model_alias == item["model_name"]:
@@ -2280,7 +2280,7 @@ async def test_proxy_model_group_info_rerank(prisma_client):  # noqa: F811  # py
     resp = await model_info_v1(
         user_api_key_dict=UserAPIKeyAuth(models=[]),
     )
-    models = resp["data"]
+    models = json.loads(resp.body)["data"]
     assert models[0]["model_info"]["mode"] == "rerank"
     resp = await model_group_info(
         user_api_key_dict=UserAPIKeyAuth(models=[]),

tests/router_unit_tests/test_router_helper_utils.py

@@ -1,3 +1,4 @@
+import json
 import os
 import traceback
 from dotenv import load_dotenv
@@ -628,17 +629,29 @@ def test_deployment_callback_respects_cooldown_time(model_list):
         assert mock_set.call_args.kwargs["time_to_cooldown"] == 0
 
 
-def test_log_retry(model_list):
-    """Test if the '_log_retry' function is working correctly"""
-    import time
-
+@pytest.mark.parametrize("metadata_key", ["metadata", "litellm_metadata"])
+def test_log_retry(model_list, metadata_key):
+    """log_retry appends one flat record per failed attempt and copies neither the request kwargs nor
+    the request metadata into it"""
     router = Router(model_list=model_list)
     new_kwargs = router.log_retry(
-        kwargs={"metadata": {}},
-        e=Exception(),
+        kwargs={
+            "model": "gpt-3.5-turbo",
+            "api_key": "sk-must-not-be-recorded",
+            "messages": [{"role": "user", "content": "hi"}],
+            metadata_key: {"model_info": {"id": "deployment-1"}, "attempted_retries": 2, "user_api_key": "sk-proxy"},
+        },
+        e=litellm.RateLimitError(message="slow down", llm_provider="openai", model="gpt-3.5-turbo"),
     )
-    assert "metadata" in new_kwargs
-    assert "previous_models" in new_kwargs["metadata"]
+    assert json.loads(json.dumps(new_kwargs[metadata_key]["previous_models"])) == [
+        {
+            "model_group": "gpt-3.5-turbo",
+            "deployment_id": "deployment-1",
+            "exception_type": "RateLimitError",
+            "exception_string": "litellm.RateLimitError: slow down",
+            "attempted_retries": 2,
+        }
+    ]
 
 
 def test_update_usage(model_list):

tests/test_litellm/caching/test_caching.py

@@ -1,9 +1,12 @@
 import logging
 import re
+from unittest.mock import MagicMock
 
 import pytest
 
+import litellm.caching.redis_cache as redis_cache_module
 from litellm.caching.caching import Cache
+from litellm.caching.redis_cache import RedisCache, _RedisTimeoutLogThrottle
 from litellm.types.caching import LiteLLMCacheType, SemanticCacheScope
 from litellm.types.utils import Embedding, EmbeddingResponse, Usage
 
@@ -53,6 +56,29 @@ def test_cache_key_debug_log_does_not_include_prompt_material(caplog):
     assert any(cache_key in message for message in created_cache_key_logs)
 
 
+@pytest.mark.parametrize(
+    ("backend", "expected_level"),
+    [
+        pytest.param(MagicMock(spec=RedisCache), logging.DEBUG, id="redis_backend_is_throttled"),
+        pytest.param(MagicMock(), logging.ERROR, id="other_backend_logs_every_timeout"),
+    ],
+)
+def test_add_cache_timeout_only_joins_redis_throttle_for_redis_backends(backend, expected_level, caplog, monkeypatch):
+    throttle = _RedisTimeoutLogThrottle(interval=5.0, clock=MagicMock(return_value=1_000.0))
+    assert throttle.admit() == 0
+    monkeypatch.setattr(redis_cache_module, "_redis_timeout_log_throttle", throttle)
+
+    cache = Cache(type=LiteLLMCacheType.LOCAL)
+    backend.set_cache.side_effect = TimeoutError("lit7520 backend timed out")
+    cache.cache = backend
+
+    with caplog.at_level(logging.DEBUG, logger="LiteLLM"):
+        cache.add_cache("result", model="gpt-4.1-mini", messages=[{"role": "user", "content": "hi"}])
+
+    records = [r for r in caplog.records if "lit7520 backend timed out" in r.getMessage()]
+    assert [r.levelno for r in records] == [expected_level]
+
+
 def _embedding_response(prompt_tokens, num_items):
     return EmbeddingResponse(
         model="amazon.titan-embed-image-v1",

tests/test_litellm/caching/test_dual_cache.py

@@ -704,3 +704,58 @@ async def test_open_breaker_keeps_async_batch_read_memory_hits_and_releases_rese
 
     assert list(await cache.async_batch_get_cache(["k1", "k2"])) == ["v1", None]
     assert "k2" not in cache.last_redis_batch_access_time
+
+
+@pytest.mark.asyncio
+async def test_redis_timeouts_falling_back_to_memory_log_once_per_interval(caplog, monkeypatch):
+    """The first fallback WARNING of a timeout streak logs, the rest stay at DEBUG until the summary."""
+    from redis.exceptions import TimeoutError as RedisTimeoutError
+
+    from litellm.caching import redis_cache as redis_cache_module
+    from litellm.caching.redis_cache import _RedisTimeoutLogThrottle
+
+    clock = MagicMock(return_value=1_000.0)
+    monkeypatch.setattr(
+        redis_cache_module, "_redis_timeout_log_throttle", _RedisTimeoutLogThrottle(interval=5.0, clock=clock)
+    )
+
+    class _TimingOutRedis:
+        async def async_increment_pipeline(self, increment_list, **kwargs):
+            raise RedisTimeoutError("Timeout reading from 127.0.0.1:6379")
+
+        async def async_increment(self, key, value, **kwargs):
+            raise RedisTimeoutError("Timeout reading from 127.0.0.1:6379")
+
+    cache = DualCache(
+        in_memory_cache=InMemoryCache(),
+        redis_cache=_TimingOutRedis(),  # pyright: ignore[reportArgumentType]  # duck-typed Redis double
+    )
+    increments = [RedisPipelineIncrementOperation(key="k", increment_value=1.0, ttl=60)]
+
+    with caplog.at_level(logging.DEBUG, logger="LiteLLM"):
+        for _ in range(100):
+            await cache.async_increment_cache_pipeline(increment_list=increments)
+            await cache.async_increment_cache("k", 1.0)
+
+    visible = [r for r in caplog.records if r.levelno >= logging.WARNING]
+    assert [(r.levelno, r.getMessage()) for r in visible] == [
+        (
+            logging.WARNING,
+            "Redis async_increment_cache_pipeline failed, falling back to in-memory result:"
+            " Timeout reading from 127.0.0.1:6379",
+        )
+    ]
+    assert visible[0].filename == "dual_cache.py"
+    assert sum("Timeout reading from" in r.getMessage() for r in caplog.records) == 200
+
+    caplog.clear()
+    clock.return_value += 5.0
+    with caplog.at_level(logging.DEBUG, logger="LiteLLM"):
+        await cache.async_increment_cache("k", 1.0)
+    assert [(r.levelno, r.getMessage()) for r in caplog.records] == [
+        (
+            logging.WARNING,
+            "Redis async_increment_cache failed, falling back to in-memory result: Timeout reading from 127.0.0.1:6379"
+            " (199 more Redis timeouts since the previous Redis timeout line were logged at DEBUG)",
+        )
+    ]

tests/test_litellm/caching/test_redis_cache.py

@@ -978,17 +978,17 @@ async def test_stale_timeout_does_not_let_sub_threshold_hard_failures_open_the_b
     from redis.exceptions import ConnectionError as RedisConnectionError
     from redis.exceptions import TimeoutError as RedisTimeoutError
 
-    from litellm.caching.redis_cache import RedisCircuitBreaker, _is_redis_timeout_failure
+    from litellm.caching.redis_cache import RedisCircuitBreaker, is_redis_timeout_failure
 
     breaker = RedisCircuitBreaker(failure_threshold=3, recovery_timeout=60, timeout_min_duration=0.05)
 
-    breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisTimeoutError("read timed out")))
+    breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisTimeoutError("read timed out")))
     await asyncio.sleep(0.06)
     for _ in range(breaker.failure_threshold - 1):
-        breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisConnectionError("refused")))
+        breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisConnectionError("refused")))
     assert breaker.is_open() is False, "2 hard failures and 1 stale timeout are below both thresholds"
 
-    breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisConnectionError("refused")))
+    breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisConnectionError("refused")))
     assert breaker.is_open() is True, "the threshold-th hard failure must still open it"
 
 
@@ -1000,19 +1000,19 @@ async def test_hard_failure_resets_timeout_streak_so_a_later_burst_must_earn_its
     from redis.exceptions import ConnectionError as RedisConnectionError
     from redis.exceptions import TimeoutError as RedisTimeoutError
 
-    from litellm.caching.redis_cache import RedisCircuitBreaker, _is_redis_timeout_failure
+    from litellm.caching.redis_cache import RedisCircuitBreaker, is_redis_timeout_failure
 
     breaker = RedisCircuitBreaker(failure_threshold=3, recovery_timeout=60, timeout_min_duration=0.05)
 
-    breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisTimeoutError("read timed out")))
-    breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisConnectionError("refused")))
+    breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisTimeoutError("read timed out")))
+    breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisConnectionError("refused")))
     await asyncio.sleep(0.06)
     for _ in range(breaker.failure_threshold):
-        breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisTimeoutError("read timed out")))
+        breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisTimeoutError("read timed out")))
     assert breaker.is_open() is False, "the burst is instantaneous, so the duration gate must hold it closed"
 
     await asyncio.sleep(0.06)
-    breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisTimeoutError("read timed out")))
+    breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisTimeoutError("read timed out")))
     assert breaker.is_open() is True, "the same run of timeouts persisting past the duration must open it"
 
 
@@ -1023,7 +1023,7 @@ async def test_breaker_metrics_track_state_and_failure_class():
     from redis.exceptions import ConnectionError as RedisConnectionError
     from redis.exceptions import TimeoutError as RedisTimeoutError
 
-    from litellm.caching.redis_cache import RedisCircuitBreaker, _is_redis_timeout_failure
+    from litellm.caching.redis_cache import RedisCircuitBreaker, is_redis_timeout_failure
 
     def sample(name, labels=None):
         return REGISTRY.get_sample_value(name, labels) or 0.0
@@ -1035,9 +1035,9 @@ async def test_breaker_metrics_track_state_and_failure_class():
     closed_gauge_before = sample("litellm_redis_circuit_breaker_state", {"state": "closed"})
 
     breaker = RedisCircuitBreaker(failure_threshold=2, recovery_timeout=60, timeout_min_duration=5.0)
-    breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisTimeoutError("t")))
-    breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisConnectionError("refused")))
-    breaker.record_failure(is_timeout=_is_redis_timeout_failure(RedisConnectionError("refused")))
+    breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisTimeoutError("t")))
+    breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisConnectionError("refused")))
+    breaker.record_failure(is_timeout=is_redis_timeout_failure(RedisConnectionError("refused")))
 
     assert sample("litellm_redis_circuit_breaker_failures_total", {"failure_class": "timeout"}) == timeout_before + 1
     assert sample("litellm_redis_circuit_breaker_failures_total", {"failure_class": "connectivity"}) == hard_before + 2
@@ -1205,6 +1205,117 @@ async def test_a_probe_overtaken_by_a_later_outage_leaves_the_breaker_to_the_new
     assert breaker._state == breaker.CLOSED
 
 
+def test_timeouts_during_a_blip_log_once_per_interval_not_once_per_call(sync_batch_redis_cache, caplog, monkeypatch):
+    """A timeout streak logs its first failure plus one summary per interval; other failures log per call."""
+    import logging
+
+    from redis.exceptions import TimeoutError as RedisTimeoutError
+
+    from litellm.caching import redis_cache as redis_cache_module
+    from litellm.caching.redis_cache import _RedisTimeoutLogThrottle
+
+    clock = MagicMock(return_value=1_000.0)
+    monkeypatch.setattr(
+        redis_cache_module, "_redis_timeout_log_throttle", _RedisTimeoutLogThrottle(interval=5.0, clock=clock)
+    )
+    sync_batch_redis_cache.redis_client.get.side_effect = RedisTimeoutError("Timeout reading from 127.0.0.1:6379")
+    sync_batch_redis_cache.redis_client.mget.side_effect = RedisTimeoutError("Timeout reading from 127.0.0.1:6379")
+
+    with caplog.at_level(logging.DEBUG, logger="LiteLLM"):
+        for _ in range(200):
+            assert sync_batch_redis_cache.get_cache("lit7520") is None
+        assert sync_batch_redis_cache.batch_get_cache(key_list=["lit7520"]) == {}
+
+    timeout_records = [r for r in caplog.records if "Timeout reading from" in r.getMessage()]
+    assert len(timeout_records) == 201, "every timeout must stay visible at DEBUG"
+    assert [r.getMessage() for r in timeout_records if r.levelno >= logging.WARNING] == [
+        "litellm.caching.caching: get() - Got exception from REDIS: Timeout reading from 127.0.0.1:6379"
+    ]
+    assert timeout_records[0].levelno == logging.ERROR
+    assert timeout_records[0].filename == "redis_cache.py"
+    assert timeout_records[0].lineno != timeout_records[-1].lineno, "the record must point at the cache operation"
+
+    caplog.clear()
+    clock.return_value += 5.0
+    with caplog.at_level(logging.DEBUG, logger="LiteLLM"):
+        assert sync_batch_redis_cache.batch_get_cache(key_list=["lit7520"]) == {}
+    assert [(r.levelno, r.getMessage()) for r in caplog.records] == [
+        (
+            logging.ERROR,
+            "Error occurred in batch get cache: Timeout reading from 127.0.0.1:6379"
+            " (200 more Redis timeouts since the previous Redis timeout line were logged at DEBUG)",
+        )
+    ]
+
+    caplog.clear()
+    sync_batch_redis_cache.redis_client.get.side_effect = OSError("redis unavailable")
+    with caplog.at_level(logging.DEBUG, logger="LiteLLM"):
+        for _ in range(3):
+            assert sync_batch_redis_cache.get_cache("lit7520") is None
+    assert [r.levelno for r in caplog.records if "redis unavailable" in r.getMessage()] == [logging.ERROR] * 3
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "call_method",
+    [
+        pytest.param(lambda c: c.async_set_cache_pipeline([("lit7520", "v")]), id="async_set_cache_pipeline"),
+        pytest.param(
+            lambda c: c.async_set_cache_pipeline_with_ttls([("lit7520", "v", 60.0)]),
+            id="async_set_cache_pipeline_with_ttls",
+        ),
+        pytest.param(lambda c: c.async_set_cache_sadd("lit7520", ["v"], ttl=None), id="async_set_cache_sadd"),
+        pytest.param(lambda c: c.async_increment("lit7520", 1.0), id="async_increment"),
+        pytest.param(
+            lambda c: c.async_increment_pipeline([{"key": "lit7520", "increment_value": 1.0, "ttl": 60}]),
+            id="async_increment_pipeline",
+        ),
+        pytest.param(lambda c: c.async_rpush("lit7520", ["v"]), id="async_rpush"),
+        pytest.param(
+            lambda c: c.async_rpush_pipeline([{"key": "lit7520", "values": ["v"]}]), id="async_rpush_pipeline"
+        ),
+        pytest.param(lambda c: c.async_lpop("lit7520"), id="async_lpop"),
+        pytest.param(lambda c: c.async_lpop_pipeline([{"key": "lit7520", "count": 1}]), id="async_lpop_pipeline"),
+    ],
+)
+async def test_write_path_timeouts_inside_the_interval_stay_at_debug(call_method, caplog, monkeypatch, redis_no_ping):
+    """A write or list operation timing out mid-streak is counted by the throttle instead of logging its own ERROR."""
+    import contextlib
+    import logging
+
+    from redis.exceptions import TimeoutError as RedisTimeoutError
+
+    from litellm.caching import redis_cache as redis_cache_module
+    from litellm.caching.redis_cache import _RedisTimeoutLogThrottle
+
+    clock = MagicMock(return_value=1_000.0)
+    throttle = _RedisTimeoutLogThrottle(interval=5.0, clock=clock)
+    assert throttle.admit() == 0
+    monkeypatch.setattr(redis_cache_module, "_redis_timeout_log_throttle", throttle)
+
+    timeout = RedisTimeoutError("Timeout reading from 127.0.0.1:6379")
+    client = MagicMock()
+    client.pipeline.return_value.__aenter__.side_effect = timeout
+    client.sadd = AsyncMock(side_effect=timeout)
+    client.incrbyfloat = AsyncMock(side_effect=timeout)
+    client.rpush = AsyncMock(side_effect=timeout)
+    client.lpop = AsyncMock(side_effect=timeout)
+    monkeypatch.setenv("REDIS_HOST", "https://my-test-host")
+    cache = RedisCache()
+
+    with (
+        patch.object(cache, "init_async_client", return_value=client),
+        caplog.at_level(logging.DEBUG, logger="LiteLLM"),
+    ):
+        with contextlib.suppress(RedisTimeoutError):
+            await call_method(cache)
+
+    timeout_records = [r for r in caplog.records if "Timeout reading from" in r.getMessage()]
+    assert [(r.levelno, r.filename) for r in timeout_records] == [(logging.DEBUG, "redis_cache.py")]
+    clock.return_value += 5.0
+    assert throttle.admit() == 1
+
+
 @pytest.mark.asyncio
 async def test_pool_wait_timeout_is_a_timeout_failure_not_hard_connectivity():
     """A saturated blocking pool must not open the breaker before the timeout minimum duration.
@@ -1241,7 +1352,7 @@ async def test_pool_wait_timeout_is_a_timeout_failure_not_hard_connectivity():
 def test_timeout_classification_follows_the_explicit_cause_chain_only():
     from redis.exceptions import ConnectionError as RedisConnectionError
 
-    from litellm.caching.redis_cache import _is_redis_timeout_failure
+    from litellm.caching.redis_cache import is_redis_timeout_failure
 
     def raise_chained_from_timeout() -> None:
         try:
@@ -1260,9 +1371,9 @@ def test_timeout_classification_follows_the_explicit_cause_chain_only():
     with pytest.raises(RedisConnectionError) as contextual:
         raise_while_handling_timeout()
 
-    assert _is_redis_timeout_failure(chained.value) is True
-    assert _is_redis_timeout_failure(contextual.value) is False
-    assert _is_redis_timeout_failure(RedisConnectionError("refused")) is False
+    assert is_redis_timeout_failure(chained.value) is True
+    assert is_redis_timeout_failure(contextual.value) is False
+    assert is_redis_timeout_failure(RedisConnectionError("refused")) is False
 
 
 class _RoundTripCountingRedis:

tests/test_litellm/integrations/otel/test_otel_v2_emitter.py

@@ -1,6 +1,8 @@
 """Golden tests for the OTel v2 engine: span shape, kinds, semconv attributes,
 legacy dual-emit, hierarchy, error status, and idempotency. Needs the OTel SDK."""
 
+import json
+
 import pytest
 
 pytest.importorskip("opentelemetry")
@@ -18,6 +20,7 @@ from litellm.integrations.otel.plumbing import providers  # noqa: E402
 from litellm.integrations.otel.emitter import SpanEmitter  # noqa: E402
 from litellm.integrations.otel.emitter import stamp_error  # noqa: E402
 from litellm.integrations.otel.mappers.utils import (  # noqa: E402
+    MAX_MESSAGE_ATTRS_PER_SPAN,
     MAX_TOOL_DEFINITION_ATTRS_PER_SPAN,
 )
 from litellm.integrations.otel.model.payloads import (  # noqa: E402
@@ -440,3 +443,161 @@ def test_vendor_tool_definitions_are_truncated_not_dropped():
     assert a["llm.tools.0.tool.name"] == "tool_0"
     assert a["llm.tools.0.tool.json_schema"]
     assert "llm.tools.126.tool.name" not in a
+
+
+def _conversation_payload(turns, choices=1, **overrides):
+    """A ``turns``-message chat with ``choices`` response choices, content-bearing."""
+    return _payload(
+        messages=[{"role": ("user", "assistant")[i % 2], "content": f"turn {i}"} for i in range(turns)],
+        response={
+            "id": "resp_1",
+            "model": "gpt-4o-2024",
+            "choices": [
+                {"finish_reason": "stop", "message": {"role": "assistant", "content": f"reply {i}"}}
+                for i in range(choices)
+            ],
+        },
+        **overrides,
+    )
+
+
+def _conversation_span(mapper_names, payload, legacy_compat=False):
+    """The exported LLM-call span for ``payload`` with content capture on."""
+    cfg = OpenTelemetryV2Config(
+        exporter="in_memory",
+        legacy_compat=legacy_compat,
+        mapper_names=list(mapper_names),
+        capture_message_content="span_only",
+    )
+    provider, exporter = providers.in_memory_provider(cfg)
+    engine = SpanEmitter(providers.get_tracer(provider, "litellm-test"), cfg)
+    engine.emit(
+        SpanRole.LLM_CALL,
+        LLMCallSpanData.from_standard_logging_payload(payload, capture_content=True),
+    )
+    (span,) = exporter.get_finished_spans()
+    return span
+
+
+def _indexed_message_count(attributes, prefix):
+    return len({key.split(".")[2] for key in attributes if key.startswith(f"{prefix}.")})
+
+
+@pytest.mark.parametrize("turns", [60, 200])
+def test_long_conversation_does_not_evict_core_attributes(turns):
+    """Per-message OpenInference attributes must never crowd core telemetry off the span."""
+    span = _conversation_span(["genai", "openinference"], _conversation_payload(turns))
+    a = span.attributes
+
+    assert span.dropped_attributes == 0
+    assert a[GenAI.REQUEST_MODEL] == "gpt-4o"
+    assert a[GenAI.PROVIDER_NAME] == "openai"
+    assert a[GenAI.USAGE_INPUT_TOKENS] == 10
+    assert a[GenAI.USAGE_OUTPUT_TOKENS] == 5
+    assert a[GenAI.RESPONSE_FINISH_REASONS] == ("stop",)
+    assert a[f"{LiteLLM.COST_PREFIX}total"] == 0.002
+
+    assert a["llm.input_messages.0.message.content"] == "turn 0"
+    assert a["llm.output_messages.0.message.content"] == "reply 0"
+    assert a[f"llm.input_messages.{turns - 1}.message.content"] == f"turn {turns - 1}"
+    assert f"llm.input_messages.{turns // 2}.message.role" not in a
+    assert len(json.loads(a["input.value"])) == turns
+    assert len(json.loads(a["output.value"])) == 1
+    assert len(json.loads(a[GenAI.INPUT_MESSAGES])) == turns
+
+
+def test_short_conversation_keeps_every_message_indexed():
+    """Below the cap nothing is truncated in either direction."""
+    a = _conversation_span(["genai", "openinference"], _conversation_payload(4, choices=2)).attributes
+    for idx in range(4):
+        assert a[f"llm.input_messages.{idx}.message.content"] == f"turn {idx}"
+    for idx in range(2):
+        assert a[f"llm.output_messages.{idx}.message.content"] == f"reply {idx}"
+
+
+def test_indexed_prompt_keeps_opener_and_latest_turns_under_a_value_length_limit(monkeypatch):
+    """The system prompt and the live turn keep their own keys once the SDK clips ``input.value``."""
+    monkeypatch.setenv("OTEL_SPAN_ATTRIBUTE_VALUE_LENGTH_LIMIT", "256")
+    chat = _conversation_payload(60)
+    payload = {
+        **chat,
+        "messages": [
+            {"role": "system", "content": "be terse"},
+            *chat["messages"][1:-1],
+            {"role": "user", "content": "LATEST-TURN"},
+        ],
+    }
+    a = _conversation_span(["genai", "openinference"], payload).attributes
+
+    assert len(a["input.value"]) == 256
+    assert a["llm.input_messages.0.message.role"] == "system"
+    assert a["llm.input_messages.0.message.content"] == "be terse"
+    assert a["llm.input_messages.59.message.role"] == "user"
+    assert a["llm.input_messages.59.message.content"] == "LATEST-TURN"
+    assert a["llm.output_messages.0.message.content"] == "reply 0"
+    assert [int(key.split(".")[2]) for key in a if key.endswith("message.content") and key.startswith("llm.input_")] == [
+        0,
+        *range(54, 60),
+    ]
+
+
+def test_message_cap_is_shared_across_input_and_output():
+    """One span-wide allowance covers both directions, and the response always keeps a share."""
+    long_prompt = _conversation_span(["genai", "openinference"], _conversation_payload(60, choices=1)).attributes
+    many_choices = _conversation_span(["genai", "openinference"], _conversation_payload(60, choices=20)).attributes
+
+    single_reply_indexed = _indexed_message_count(long_prompt, "llm.output_messages")
+    assert single_reply_indexed == 1
+    assert _indexed_message_count(long_prompt, "llm.input_messages") + single_reply_indexed == (
+        MAX_MESSAGE_ATTRS_PER_SPAN // 2
+    )
+
+    assert _indexed_message_count(many_choices, "llm.input_messages") > 0
+    assert _indexed_message_count(many_choices, "llm.output_messages") > single_reply_indexed
+    assert _indexed_message_count(many_choices, "llm.input_messages") + _indexed_message_count(
+        many_choices, "llm.output_messages"
+    ) == (MAX_MESSAGE_ATTRS_PER_SPAN // 2)
+
+
+def test_fully_populated_span_with_every_vocabulary_stays_within_the_attribute_limit():
+    """Every capped family maxed at once still leaves the whole core intact."""
+    payload = _conversation_payload(
+        200,
+        choices=20,
+        stream=True,
+        model_parameters={
+            **_tools_payload(127)["model_parameters"],
+            "top_p": 0.9,
+            "frequency_penalty": 0.1,
+            "presence_penalty": 0.1,
+            "seed": 7,
+            "stop": ["\n"],
+        },
+        cost_breakdown={
+            key: 0.001
+            for key in (
+                "input_cost",
+                "output_cost",
+                "cache_read_cost",
+                "cache_creation_cost",
+                "tool_usage_cost",
+                "original_cost",
+                "discount_amount",
+                "discount_percent",
+                "margin_fixed_amount",
+                "margin_percent",
+                "margin_total_amount",
+                "total_cost",
+            )
+        },
+    )
+    span = _conversation_span(["genai", "openinference", "langfuse", "weave", "langtrace"], payload, legacy_compat=True)
+    a = span.attributes
+
+    assert span.dropped_attributes == 0
+    assert a[GenAI.REQUEST_MODEL] == "gpt-4o"
+    assert a[f"{LiteLLM.COST_PREFIX}total"] == 0.002
+    assert a[LiteLLM.TOOLS_DECLARED] == 127
+    assert a["llm.input_messages.0.message.content"] == "turn 0"
+    assert a["llm.input_messages.199.message.content"] == "turn 199"
+    assert a["llm.output_messages.0.message.content"] == "reply 0"

tests/test_litellm/integrations/test_prometheus_labels.py

@@ -716,6 +716,77 @@ async def test_failure_hook_emits_api_provider_value_on_failed_requests_metric()
         _clear_prometheus_registry()
 
 
+async def _failed_requests_api_provider_labels(
+    request_data: dict[str, object],
+    original_exception: Exception,
+) -> list[str]:
+    from litellm.integrations.prometheus import PrometheusLogger
+    from litellm.proxy._types import UserAPIKeyAuth
+
+    _clear_prometheus_registry()
+    try:
+        await PrometheusLogger().async_post_call_failure_hook(
+            request_data=request_data,
+            original_exception=original_exception,
+            user_api_key_dict=UserAPIKeyAuth(token="tok"),
+        )
+        return [
+            s.labels.get("api_provider")
+            for s in _collected_samples("litellm_proxy_failed_requests_metric_total")
+        ]
+    finally:
+        _clear_prometheus_registry()
+
+
+@pytest.mark.asyncio
+async def test_failure_hook_emits_api_provider_from_pre_call_rate_limit_error_for_router_alias():
+    """
+    Pre-call limiters reject before a deployment lands on request_data and a
+    router alias cannot be inferred from its name, so the provider the limiter
+    resolved onto the exception is the only source for the label.
+    """
+    from litellm.exceptions import RateLimitType
+    from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitError
+
+    err = ProxyRateLimitError(
+        detail={"error": "rpm exceeded"},
+        rate_limit_type=RateLimitType.REQUESTS,
+        model="openai/gpt-5.4-mini",
+        llm_provider="openai",
+    )
+
+    assert await _failed_requests_api_provider_labels(
+        {"model": "team-chat-model", "metadata": {}}, err
+    ) == ["openai"]
+
+
+@pytest.mark.asyncio
+async def test_failure_hook_leaves_api_provider_unset_when_rate_limiter_could_not_resolve_provider():
+    from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitError
+
+    err = ProxyRateLimitError(detail={"error": "rpm exceeded"}, model="unknown-alias")
+
+    assert await _failed_requests_api_provider_labels(
+        {"model": "unknown-alias", "metadata": {}}, err
+    ) == ["None"]
+
+
+@pytest.mark.asyncio
+async def test_failure_hook_prefers_request_data_provider_over_exception_provider():
+    from litellm.exceptions import RateLimitError
+
+    err = RateLimitError(message="upstream 429", llm_provider="openai", model="gpt-4o")
+
+    assert await _failed_requests_api_provider_labels(
+        {
+            "model": "gpt-4o",
+            "metadata": {},
+            "litellm_params": {"custom_llm_provider": "azure"},
+        },
+        err,
+    ) == ["azure"]
+
+
 if __name__ == "__main__":
     test_user_email_in_required_metrics()
     test_user_email_label_exists()

tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py

@@ -2101,6 +2101,37 @@ def test_generic_cost_per_token_azure_gpt_6_astra_foundry_price_sheet(
     assert completion_cost == pytest.approx(zone_multiplier * output_multiplier * completion_tokens * 5e-5)
 
 
+@pytest.mark.parametrize(
+    "model,input_rate,cache_read_rate,output_rate",
+    [
+        ("azure/gpt-chat-latest", 5e-6, 5e-7, 3e-5),
+        ("azure/chat-latest", 5e-6, 5e-7, 3e-5),
+        ("azure/us/gpt-chat-latest", 5.5e-6, 5.5e-7, 3.3e-5),
+    ],
+)
+def test_generic_cost_per_token_azure_gpt_chat_latest_price_sheet(
+    _local_model_cost_map, model, input_rate, cache_read_rate, output_rate
+):
+    """The Azure OpenAI price sheet lists GPT-Chat Latest at $5 input, $0.50 cached input and $30 output per 1M
+    tokens on Global, and $5.50, $0.55 and $33 on Data Zone. Foundry names the product gpt-chat-latest and the
+    OpenAI API names the same model chat-latest, so both spellings bill the Global sheet.
+    """
+    prompt_tokens = 100000
+    cached_tokens = 40000
+    completion_tokens = 1000
+    usage = Usage(
+        prompt_tokens=prompt_tokens,
+        completion_tokens=completion_tokens,
+        total_tokens=prompt_tokens + completion_tokens,
+        prompt_tokens_details=PromptTokensDetailsWrapper(cached_tokens=cached_tokens),
+    )
+
+    prompt_cost, completion_cost = generic_cost_per_token(model=model, usage=usage, custom_llm_provider="azure")
+
+    assert prompt_cost == pytest.approx((prompt_tokens - cached_tokens) * input_rate + cached_tokens * cache_read_rate)
+    assert completion_cost == pytest.approx(completion_tokens * output_rate)
+
+
 def test_generic_cost_per_token_azure_ai_gpt_6_astra_flex_bills_the_standard_rate(_local_model_cost_map):
     usage = Usage(prompt_tokens=1000, completion_tokens=100, total_tokens=1100)
 
@@ -2648,6 +2679,7 @@ def test_cache_writing_cost_with_zero_creation_tokens_and_ephemeral_details():
 
     prompt_tokens_details: PromptTokensDetailsResult = {
         "cache_hit_tokens": 0,
+        "cache_hit_audio_tokens": 0,
         "cache_creation_tokens": 0,
         "cache_creation_token_details": CacheCreationTokenDetails(
             ephemeral_5m_input_tokens=100,
@@ -4005,6 +4037,7 @@ def test_billed_token_rates_follow_the_token_tier_the_breakdown_bills_at(monkeyp
         input_cost_per_token=6e-6,
         output_cost_per_token=3e-5,
         cache_read_input_token_cost=6e-7,
+        cache_read_input_audio_token_cost=6e-7,
         cache_creation_input_token_cost=7.5e-6,
         cache_creation_input_token_cost_above_1hr=0.0,
         output_cost_per_reasoning_token=3e-5,
@@ -5147,3 +5180,157 @@ def test_generic_cost_per_token_bills_nested_reasoning_once_beside_audio_output(
     assert completion_cost == pytest.approx(
         30 * info["output_cost_per_token"] + 70 * info["output_cost_per_audio_token"]
     )
+
+
+def test_cached_realtime_audio_tokens_billed_at_audio_cache_read_rate(
+    _local_model_cost_map: None,
+) -> None:
+    usage = Usage(
+        prompt_tokens=283,
+        completion_tokens=0,
+        total_tokens=283,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            text_tokens=116,
+            audio_tokens=167,
+            cached_tokens=192,
+            cached_tokens_details={"text_tokens": 64, "audio_tokens": 128},
+        ),
+    )
+
+    prompt_cost, _ = generic_cost_per_token(
+        model="gpt-realtime-2", usage=usage, custom_llm_provider="openai"
+    )
+    assert prompt_cost == pytest.approx(0.0015328)
+
+
+def test_prompt_tokens_details_without_cached_tokens_details_unchanged(
+    _local_model_cost_map: None,
+) -> None:
+    usage = Usage(
+        prompt_tokens=283,
+        completion_tokens=0,
+        total_tokens=283,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            text_tokens=116, audio_tokens=167, cached_tokens=192
+        ),
+    )
+
+    prompt_cost, _ = generic_cost_per_token(
+        model="gpt-realtime-2", usage=usage, custom_llm_provider="openai"
+    )
+    assert prompt_cost == pytest.approx(0.0029888)
+
+
+def test_cached_audio_tokens_fall_back_to_cache_read_input_token_cost() -> None:
+    model_info: ModelInfo = {
+        "input_cost_per_token": 4e-6,
+        "input_cost_per_audio_token": 32e-6,
+        "cache_read_input_token_cost": 5e-7,
+    }
+    usage = Usage(
+        prompt_tokens=283,
+        completion_tokens=0,
+        total_tokens=283,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            text_tokens=116,
+            audio_tokens=167,
+            cached_tokens=192,
+            cached_tokens_details={"text_tokens": 64, "audio_tokens": 128},
+        ),
+    )
+
+    prompt_cost, _ = generic_cost_per_token(
+        model="some-realtime-model",
+        usage=usage,
+        custom_llm_provider="openai",
+        model_info=model_info,
+    )
+    expected = 52 * 4e-6 + 64 * 5e-7 + 39 * 32e-6 + 128 * 5e-7
+    assert prompt_cost == pytest.approx(expected)
+
+
+def test_cached_audio_tokens_capped_at_cached_tokens(_local_model_cost_map: None) -> None:
+    """Nested cached_tokens_details exceeding cached_tokens must not over-subtract the audio bucket."""
+    usage = Usage(
+        prompt_tokens=283,
+        completion_tokens=0,
+        total_tokens=283,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            text_tokens=116,
+            audio_tokens=167,
+            cached_tokens=100,
+            cached_tokens_details={"audio_tokens": 128},
+        ),
+    )
+
+    prompt_cost, _ = generic_cost_per_token(
+        model="gpt-realtime-2", usage=usage, custom_llm_provider="openai"
+    )
+    assert prompt_cost == pytest.approx(116 * 4e-6 + (167 - 100) * 32e-6 + 100 * 4e-7)
+
+
+def test_cached_audio_tokens_billed_at_audio_cache_rate_through_model_info_lookup(_local_model_cost_map: None) -> None:
+    usage = Usage(
+        prompt_tokens=1000,
+        completion_tokens=0,
+        total_tokens=1000,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            text_tokens=400,
+            audio_tokens=600,
+            cached_tokens=500,
+            cached_tokens_details={"text_tokens": 100, "audio_tokens": 400},
+        ),
+    )
+
+    prompt_cost, _ = generic_cost_per_token(model="gpt-realtime-2.1-mini", usage=usage, custom_llm_provider="openai")
+    assert prompt_cost == pytest.approx(300 * 6e-7 + 100 * 6e-8 + 200 * 1e-5 + 400 * 3e-7)
+
+
+def test_cache_read_breakdown_splits_cached_audio_at_the_audio_cache_rate(_local_model_cost_map: None) -> None:
+    usage = Usage(
+        prompt_tokens=4863,
+        completion_tokens=1087,
+        total_tokens=5950,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            text_tokens=1693,
+            audio_tokens=3170,
+            cached_tokens=2816,
+            cached_tokens_details={"text_tokens": 896, "audio_tokens": 1920},
+        ),
+    )
+
+    breakdown = get_token_type_cost_breakdown(model="gpt-realtime-2.1-mini", custom_llm_provider="openai", usage=usage)
+    prompt_cost, _ = generic_cost_per_token(model="gpt-realtime-2.1-mini", usage=usage, custom_llm_provider="openai")
+
+    assert breakdown.cache_read_cost == pytest.approx(896 * 6e-8 + 1920 * 3e-7)
+    assert breakdown.rates is not None
+    assert breakdown.rates.cache_read_input_audio_token_cost == pytest.approx(3e-7)
+    assert prompt_cost == pytest.approx((1693 - 896) * 6e-7 + (3170 - 1920) * 1e-5 + breakdown.cache_read_cost)
+
+
+@pytest.mark.parametrize(
+    ("model", "custom_llm_provider", "expected_prompt_cost"),
+    (
+        pytest.param("azure/gpt-realtime-2025-08-28", "azure", 300 * 4e-6 + 100 * 4e-7 + 200 * 3.2e-5 + 400 * 4e-7, id="azure-gpt-realtime"),
+        pytest.param("azure/gpt-realtime-1.5-2026-02-23", "azure", 300 * 4e-6 + 100 * 4e-7 + 200 * 3.2e-5 + 400 * 4e-7, id="azure-gpt-realtime-1.5"),
+        pytest.param("azure/gpt-realtime-mini", "azure", 300 * 6e-7 + 100 * 6e-8 + 200 * 1e-5 + 400 * 3e-7, id="azure-gpt-realtime-mini"),
+        pytest.param("gpt-realtime-mini", "openai", 300 * 6e-7 + 100 * 6e-8 + 200 * 1e-5 + 400 * 3e-7, id="openai-gpt-realtime-mini"),
+    ),
+)
+def test_realtime_models_bill_cached_text_and_audio_at_their_cache_read_rates(
+    _local_model_cost_map: None, model: str, custom_llm_provider: str, expected_prompt_cost: float
+) -> None:
+    usage = Usage(
+        prompt_tokens=1000,
+        completion_tokens=0,
+        total_tokens=1000,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            text_tokens=400,
+            audio_tokens=600,
+            cached_tokens=500,
+            cached_tokens_details={"text_tokens": 100, "audio_tokens": 400},
+        ),
+    )
+
+    prompt_cost, _ = generic_cost_per_token(model=model, usage=usage, custom_llm_provider=custom_llm_provider)
+    assert prompt_cost == pytest.approx(expected_prompt_cost)

tests/test_litellm/litellm_core_utils/test_fallback_generalizations.py

@@ -11,12 +11,12 @@ import logging
 
 import pytest
 
-
 import litellm
 from litellm._logging import verbose_logger
 from litellm.litellm_core_utils.fallback_generalizations import (
     get_fallback_generalization_rules,
     match_capability_generalizations,
+    match_fill_missing_generalizations,
     match_routing_generalization,
     set_fallback_generalizations,
 )
@@ -116,6 +116,67 @@ def test_capability_union_is_last_wins_in_file_order(restore_generalizations):
     }
 
 
+def test_fill_missing_requires_per_rule_opt_in(restore_generalizations):
+    restore_generalizations(
+        [
+            {"name": "base", "pattern": r"^acme-", "model_info": {"supports_reasoning": True}},
+            {
+                "name": "opt-in",
+                "pattern": r"^acme-",
+                "fill_missing_for_providers": ["openai"],
+                "model_info": {"supports_vision": True},
+            },
+        ]
+    )
+    assert match_fill_missing_generalizations("acme-1", "openai") == {"supports_vision": True}
+    assert match_fill_missing_generalizations("acme-1", "azure") is None
+    assert match_capability_generalizations("acme-1") == {
+        "supports_reasoning": True,
+        "supports_vision": True,
+    }
+
+    restore_generalizations(
+        [{"name": "base", "pattern": r"^acme-", "model_info": {"supports_reasoning": True}}]
+    )
+    assert match_fill_missing_generalizations("acme-1", "openai") is None
+
+    restore_generalizations(
+        [
+            {
+                "name": "mixed",
+                "pattern": r"^acme-",
+                "fill_missing_for_providers": ["openai"],
+                "model_info": {"litellm_provider": "openai", "supports_vision": True},
+            }
+        ]
+    )
+    assert match_fill_missing_generalizations("acme-1", "openai") == {"supports_vision": True}
+
+    restore_generalizations(
+        [
+            {
+                "name": "route",
+                "pattern": r"^acme-",
+                "fill_missing_for_providers": ["openai"],
+                "model_info": {"litellm_provider": "openai"},
+            }
+        ]
+    )
+    assert match_fill_missing_generalizations("acme-1", "openai") is None
+
+    restore_generalizations(
+        [
+            {
+                "name": "malformed",
+                "pattern": r"^acme-",
+                "fill_missing_for_providers": "openai",
+                "model_info": {"supports_vision": True},
+            }
+        ]
+    )
+    assert match_fill_missing_generalizations("acme-1", "openai") is None
+
+
 def test_routing_rules_are_excluded_from_capability_results(restore_generalizations):
     restore_generalizations(
         [
@@ -299,6 +360,76 @@ def test_exact_entry_takes_precedence_over_rule(restore_generalizations):
     assert info["input_cost_per_token"] != 999.0
 
 
+def test_exact_entries_fill_only_missing_fields(restore_generalizations, monkeypatch):
+    monkeypatch.setattr(
+        litellm,
+        "model_cost",
+        {
+            **litellm.model_cost,
+            "acme-full": {
+                "input_cost_per_token": 1e-6,
+                "output_cost_per_token": 2e-6,
+                "litellm_provider": "openai",
+                "mode": "chat",
+                "max_tokens": 7,
+                "supports_reasoning": False,
+            },
+            "acme-bare": {
+                "input_cost_per_token": 3e-6,
+                "output_cost_per_token": 4e-6,
+                "litellm_provider": "openai",
+                "mode": "chat",
+            },
+            "acme-image": {
+                "input_cost_per_token": 5e-6,
+                "output_cost_per_token": 6e-6,
+                "litellm_provider": "openai",
+                "mode": "image_generation",
+            },
+            "acme-other": {
+                "input_cost_per_token": 7e-6,
+                "output_cost_per_token": 8e-6,
+                "litellm_provider": "openrouter",
+                "mode": "chat",
+            },
+        },
+    )
+    restore_generalizations(
+        [
+            {
+                "name": "acme-backfill",
+                "pattern": r"^acme-",
+                "fill_missing_for_providers": ["openai"],
+                "model_info": {"supports_reasoning": True, "max_tokens": 5},
+            }
+        ]
+    )
+    litellm.get_model_info.cache_clear()
+
+    full = litellm.get_model_info("acme-full", custom_llm_provider="openai")
+    assert full["supports_reasoning"] is False
+    assert full["max_tokens"] == 7
+
+    bare = litellm.get_model_info("acme-bare", custom_llm_provider="openai")
+    assert bare["supports_reasoning"] is True
+    assert bare["max_tokens"] == 5
+    assert bare["input_cost_per_token"] == 3e-6
+    assert bare["key"] == "acme-bare"
+
+    other = litellm.get_model_info("acme-other", custom_llm_provider="openrouter")
+    assert other.get("supports_reasoning") is None
+
+    image = litellm.get_model_info("acme-image", custom_llm_provider="openai")
+    assert image.get("supports_reasoning") is None
+
+    restore_generalizations(
+        [{"name": "acme-backfill", "pattern": r"^acme-", "model_info": {"supports_reasoning": True, "max_tokens": 5}}]
+    )
+    litellm.get_model_info.cache_clear()
+    unflagged = litellm.get_model_info("acme-bare", custom_llm_provider="openai")
+    assert unflagged.get("supports_reasoning") is None
+
+
 # --------------------------------------------------------------------------- #
 # Shipped rules (bundled cost map)
 # --------------------------------------------------------------------------- #
@@ -415,6 +546,18 @@ def test_shipped_version_boundaries(shipped_cost_map, model, provider, adaptive,
     assert info.get("supports_mid_conversation_system") is mid_conversation, model
 
 
+def test_shipped_claude_version_regex_excludes_undelimited_41(shipped_cost_map):
+    unmatched = match_capability_generalizations("github_copilot/claude-opus-41")
+    assert unmatched is None or "supports_adaptive_thinking" not in unmatched
+    assert unmatched is None or "supports_mid_conversation_system" not in unmatched
+
+    for model in ("claude-opus-5", "claude-sonnet-4-8"):
+        matched = match_capability_generalizations(model)
+        assert matched is not None
+        assert matched["supports_adaptive_thinking"] is True
+        assert matched["supports_mid_conversation_system"] is True
+
+
 def test_shipped_rules_cover_new_families_like_fable_at_5_plus(shipped_cost_map):
     """Both version gates accept any claude-<family>- id at major 5 or higher, bare
     major or major-minor, so a new family shaped like claude-fable-5 gets adaptive
@@ -605,6 +748,10 @@ def test_shipped_wandb_rule_loses_to_mapped_non_reasoning_entries(shipped_cost_m
         assert litellm.supports_reasoning(model=model, custom_llm_provider="wandb") is False, model
 
 
+def test_shipped_wandb_rule_does_not_fill_missing_mapped_entries(shipped_cost_map):
+    assert match_fill_missing_generalizations("wandb/meta-llama/Llama-3.1-8B-Instruct", "wandb") is None
+
+
 def test_shipped_wandb_rule_is_anchored_to_the_wandb_namespace(shipped_cost_map):
     """``^wandb/`` is anchored, so it cannot leak onto another provider's ids."""
     assert match_capability_generalizations("wandb/some-new-model") == {"supports_reasoning": True}
@@ -722,3 +869,58 @@ def test_shipped_openai_reasoning_rule_skips_non_reasoning_gpt_ids(shipped_cost_
 def test_shipped_openai_reasoning_rule_loses_to_mapped_entries(shipped_cost_map):
     assert "gpt-5-search-api" in litellm.model_cost
     assert litellm.supports_reasoning(model="gpt-5-search-api", custom_llm_provider="openai") is False
+
+
+@pytest.mark.parametrize(
+    "model,provider,expected_supports_reasoning",
+    [
+        ("azure/us/o1-2024-12-17", "azure", True),
+        ("github_copilot/gpt-5", "github_copilot", None),
+        ("openrouter/openai/o1", "openrouter", None),
+        ("perplexity/openai/gpt-5.4-mini", "perplexity", None),
+    ],
+)
+def test_shipped_openai_reasoning_rule_backfills_only_approved_providers(
+    shipped_cost_map, model, provider, expected_supports_reasoning
+):
+    assert model in litellm.model_cost
+    raw_entry = litellm.model_cost[model]
+    assert "supports_reasoning" not in raw_entry
+    model_without_provider = model.removeprefix(f"{provider}/")
+    info = litellm.get_model_info(model=model_without_provider, custom_llm_provider=provider)
+    assert info.get("supports_reasoning") is expected_supports_reasoning
+    assert info["input_cost_per_token"] == raw_entry.get("input_cost_per_token", 0)
+
+
+def test_shipped_openai_reasoning_rule_matches_only_openai(shipped_cost_map):
+    assert match_fill_missing_generalizations("gpt-5.4", "openai") == {"supports_reasoning": True}
+    assert match_fill_missing_generalizations("gpt-5.4", "openrouter") is None
+
+
+def test_shipped_openai_reasoning_rule_skips_non_text_modes(shipped_cost_map):
+    model = "gemini/deep-research-pro-preview-12-2025"
+    assert model in litellm.model_cost
+    raw_entry = litellm.model_cost[model]
+    assert "supports_reasoning" not in raw_entry
+    assert raw_entry["mode"] == "image_generation"
+
+    info = litellm.get_model_info("deep-research-pro-preview-12-2025", custom_llm_provider="gemini")
+    assert info.get("supports_reasoning") is None
+
+
+def test_shipped_claude_thinking_rules_backfill_only_anthropic(shipped_cost_map):
+    model = "perplexity/anthropic/claude-sonnet-4-6"
+    assert model in litellm.model_cost
+    raw_entry = litellm.model_cost[model]
+    assert "supports_adaptive_thinking" not in raw_entry
+    assert "max_input_tokens" not in raw_entry
+
+    info = litellm.get_model_info(model="anthropic/claude-sonnet-4-6", custom_llm_provider="perplexity")
+    assert info.get("supports_adaptive_thinking") is None
+    assert info.get("supports_legacy_thinking") is None
+    assert info.get("max_input_tokens") is None
+    assert match_fill_missing_generalizations("claude-sonnet-4-6", "anthropic") == {
+        "supports_adaptive_thinking": True,
+        "supports_legacy_thinking": True,
+    }
+    assert match_fill_missing_generalizations("claude-sonnet-4-6", "perplexity") is None

tests/test_litellm/litellm_core_utils/test_litellm_logging.py

@@ -5221,10 +5221,11 @@ def test_handle_anthropic_messages_parsed_response_logging_preserves_fast_mode_s
     assert getattr(result.usage, "speed", None) == "fast"
 
 
-def test_logging_init_sets_trace_id():
+def test_logging_init_sets_trace_id(monkeypatch):
     """Logging.__init__() must call set_trace_id with self.litellm_trace_id."""
     from litellm.litellm_core_utils.litellm_logging import Logging
 
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     trace_id_var.set("")
 
     log_obj = Logging(
@@ -5240,7 +5241,7 @@ def test_logging_init_sets_trace_id():
     assert trace_id_var.get() == log_obj.litellm_trace_id
 
 
-def test_logging_init_skips_stamping_when_correlation_logging_unsupported():
+def test_logging_init_skips_stamping_when_correlation_logging_unsupported(monkeypatch):
     """supports_correlation_logging=False (what wrapper(), the sync entry
     point, always passes) must leave trace_id_var/session_id_var completely
     untouched, even though self.litellm_trace_id/litellm_session_id (the
@@ -5248,6 +5249,7 @@ def test_logging_init_skips_stamping_when_correlation_logging_unsupported():
     usual - only the ambient contextvar stamping is gated."""
     from litellm.litellm_core_utils.litellm_logging import Logging
 
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     trace_id_var.set("")
     session_id_var.set("")
 
@@ -5271,10 +5273,48 @@ def test_logging_init_skips_stamping_when_correlation_logging_unsupported():
     assert log_obj.litellm_session_id == "should-not-be-stamped"
 
 
-def test_logging_init_sets_session_id_when_provided():
+def test_logging_init_skips_stamping_when_request_correlation_in_logs_disabled(monkeypatch):
+    from litellm.litellm_core_utils.litellm_logging import Logging
+
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", False)
+    trace_id_var.set("outer")
+    session_id_var.set("outer-sid")
+    try:
+        with (
+            patch(  # test-quality-ok: regression test verifies disabled stamping skips both setters
+                "litellm.litellm_core_utils.litellm_logging.set_trace_id"
+            ) as mock_set_trace_id,
+            patch(  # test-quality-ok: regression test verifies disabled stamping skips both setters
+                "litellm.litellm_core_utils.litellm_logging.set_session_id"
+            ) as mock_set_session_id,
+        ):
+            log_obj = Logging(
+                model="gpt-3.5-turbo",
+                messages=[{"role": "user", "content": "hi"}],
+                stream=False,
+                call_type="completion",
+                start_time=None,
+                litellm_call_id="call-disabled",
+                function_id="fn-disabled",
+                kwargs={"litellm_session_id": "disabled-session"},
+                supports_correlation_logging=True,
+            )
+
+            assert trace_id_var.get() == "outer"
+            assert session_id_var.get() == "outer-sid"
+            assert log_obj._own_trace_id == "outer"
+            mock_set_trace_id.assert_not_called()
+            mock_set_session_id.assert_not_called()
+    finally:
+        trace_id_var.set("")
+        session_id_var.set("")
+
+
+def test_logging_init_sets_session_id_when_provided(monkeypatch):
     """Logging.__init__() must call set_session_id when litellm_session_id is in kwargs."""
     from litellm.litellm_core_utils.litellm_logging import Logging
 
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     session_id_var.set("")
 
     Logging(
@@ -5290,11 +5330,12 @@ def test_logging_init_sets_session_id_when_provided():
     assert session_id_var.get() == "my-session-99"
 
 
-def test_logging_init_resets_session_id_to_empty_when_absent():
+def test_logging_init_resets_session_id_to_empty_when_absent(monkeypatch):
     """When no session_id is in kwargs, Logging.__init__() must reset session_id_var to ""
     so a prior request's session_id does not leak into subsequent log records."""
     from litellm.litellm_core_utils.litellm_logging import Logging
 
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     session_id_var.set("preexisting-sid")
 
     Logging(
@@ -5310,7 +5351,7 @@ def test_logging_init_resets_session_id_to_empty_when_absent():
     assert session_id_var.get() == ""
 
 
-def test_restore_correlation_context_resets_to_pre_call_value():
+def test_restore_correlation_context_resets_to_pre_call_value(monkeypatch):
     """_restore_correlation_context() must put trace_id_var/session_id_var back to
     whatever they were immediately before this Logging instance was constructed.
     This is the mechanism that prevents a nested call (e.g. a guardrail's own
@@ -5318,6 +5359,7 @@ def test_restore_correlation_context_resets_to_pre_call_value():
     session_id into the outer call's subsequent log lines."""
     from litellm.litellm_core_utils.litellm_logging import Logging
 
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     trace_id_var.set("outer-trace")
     session_id_var.set("outer-session")
     try:
@@ -5343,7 +5385,7 @@ def test_restore_correlation_context_resets_to_pre_call_value():
         session_id_var.set("")
 
 
-def test_restore_correlation_context_safe_to_call_repeatedly():
+def test_restore_correlation_context_safe_to_call_repeatedly(monkeypatch):
     """Calling _restore_correlation_context() more than once must not raise.
 
     It's deliberately NOT guarded against repeat calls: wrapper()'s finally
@@ -5353,6 +5395,7 @@ def test_restore_correlation_context_safe_to_call_repeatedly():
     the contextvars, so repeat calls are expected, not just tolerated."""
     from litellm.litellm_core_utils.litellm_logging import Logging
 
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     log_obj = Logging(
         model="gpt-3.5-turbo",
         messages=[{"role": "user", "content": "hi"}],
@@ -5367,8 +5410,40 @@ def test_restore_correlation_context_safe_to_call_repeatedly():
     log_obj._restore_correlation_context()  # must not raise
 
 
+def test_restore_correlation_context_does_not_resanitize(monkeypatch):
+    from litellm.litellm_core_utils.litellm_logging import Logging
+    from litellm._logging import _sanitize_correlation_id
+
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
+    trace_id_var.set("outer-trace")
+    session_id_var.set("outer-session")
+    try:
+        log_obj = Logging(
+            model="gpt-3.5-turbo",
+            messages=[{"role": "user", "content": "hi"}],
+            stream=False,
+            call_type="completion",
+            start_time=None,
+            litellm_call_id="call-no-resanitize",
+            function_id="fn-no-resanitize",
+            kwargs={"litellm_session_id": "inner-session"},
+        )
+
+        with patch(  # test-quality-ok: regression test verifies restore avoids sanitization
+            "litellm._logging._sanitize_correlation_id", wraps=_sanitize_correlation_id
+        ) as mock_sanitize:
+            log_obj._restore_correlation_context()
+
+        mock_sanitize.assert_not_called()
+        assert trace_id_var.get() == "outer-trace"
+        assert session_id_var.get() == "outer-session"
+    finally:
+        trace_id_var.set("")
+        session_id_var.set("")
+
+
 @pytest.mark.asyncio
-async def test_restore_correlation_context_works_across_asyncio_task_boundary():
+async def test_restore_correlation_context_works_across_asyncio_task_boundary(monkeypatch):
     """_restore_correlation_context() must succeed even when it's called from a
     different asyncio Task than the one Logging.__init__() ran in - exactly what
     happens on litellm's real async success path, where async_success_handler is
@@ -5385,6 +5460,7 @@ async def test_restore_correlation_context_works_across_asyncio_task_boundary():
     """
     from litellm.litellm_core_utils.litellm_logging import Logging
 
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     trace_id_var.set("outer-trace-cross-task")
     session_id_var.set("outer-session-cross-task")
     try:

tests/test_litellm/litellm_core_utils/test_streaming_handler.py

@@ -4100,7 +4100,7 @@ async def test_async_streaming_completion_does_not_reset_context_before_iteratio
         session_id_var.set("")
 
 
-def test_stream_wrapper_del_restores_correlation_context():
+def test_stream_wrapper_del_restores_correlation_context(monkeypatch):
     """CustomStreamWrapper.__del__ is the best-effort fallback for an abandoned
     stream (caller never exhausts it, so the normal terminal-handler restore
     never fires). Testing this via real garbage collection is unreliable in
@@ -4112,6 +4112,7 @@ def test_stream_wrapper_del_restores_correlation_context():
     doesn't run actual finalization, and this exercises exactly the logic that
     real garbage collection would eventually trigger.
     """
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     trace_id_var.set("outer-trace-abandoned")
     session_id_var.set("outer-session-abandoned")
     try:
@@ -4159,12 +4160,13 @@ def test_stream_wrapper_del_never_raises_with_broken_logging_obj():
     wrapper.__del__()  # must not raise
 
 
-def test_stream_wrapper_del_does_not_clobber_a_newer_active_call():
+def test_stream_wrapper_del_does_not_clobber_a_newer_active_call(monkeypatch):
     """A delayed finalizer must never stomp a different, still-active call's
     context. If an abandoned stream's __del__ fires late - after a new call
     has already started in the same Task/thread and claimed the contextvars -
     unconditionally restoring the abandoned stream's own pre-call snapshot
     would corrupt the active call's subsequent log lines with stale ids."""
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     trace_id_var.set("outer-trace-before-abandoned-call")
     session_id_var.set("outer-session-before-abandoned-call")
     try:
@@ -4210,13 +4212,14 @@ def test_stream_wrapper_del_does_not_clobber_a_newer_active_call():
         session_id_var.set("")
 
 
-def test_stream_wrapper_del_restores_when_own_session_id_needed_sanitizing():
+def test_stream_wrapper_del_restores_when_own_session_id_needed_sanitizing(monkeypatch):
     """The __del__ guard must compare against the *sanitized* id actually
     stored in the contextvar, not the raw litellm_session_id/litellm_trace_id
     - set_session_id()/set_trace_id() strip control characters before
     storing, so a caller-supplied id containing e.g. a newline would never
     equal the raw attribute, and the guard would wrongly conclude some other
     call has claimed the context and skip cleanup forever."""
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     trace_id_var.set("outer-trace-needs-sanitizing")
     session_id_var.set("outer-session-needs-sanitizing")
     try:
@@ -4250,7 +4253,7 @@ def test_stream_wrapper_del_restores_when_own_session_id_needed_sanitizing():
         session_id_var.set("")
 
 
-def test_stream_wrapper_next_keeps_context_active_through_synthesized_finish_reason_chunk():
+def test_stream_wrapper_next_keeps_context_active_through_synthesized_finish_reason_chunk(monkeypatch):
     """When the underlying stream ends without ever emitting an explicit
     finish_reason chunk, __next__ synthesizes one via finish_reason_handler()
     and returns it. That chunk is still this call's own data - the caller's
@@ -4261,6 +4264,7 @@ def test_stream_wrapper_next_keeps_context_active_through_synthesized_finish_rea
     correct, deterministic restore on the very next __next__() call, since
     completion_stream is already exhausted and immediately re-raises
     StopIteration."""
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     trace_id_var.set("outer-trace-finish-reason")
     session_id_var.set("outer-session-finish-reason")
     try:
@@ -4300,12 +4304,13 @@ def test_stream_wrapper_next_keeps_context_active_through_synthesized_finish_rea
         session_id_var.set("")
 
 
-def test_stream_wrapper_del_cleans_up_after_synthesized_finish_reason_chunk():
+def test_stream_wrapper_del_cleans_up_after_synthesized_finish_reason_chunk(monkeypatch):
     """A caller that breaks immediately after seeing finish_reason (the
     early-break pattern) never triggers the next()-driven restore above - it
     relies on the best-effort __del__ guard instead, same as any other
     abandoned stream. The guard must still recognize this call's own
     (unrestored) ids as unclaimed and clean them up."""
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     trace_id_var.set("outer-trace-finish-reason-del")
     session_id_var.set("outer-session-finish-reason-del")
     try:
@@ -4338,10 +4343,11 @@ def test_stream_wrapper_del_cleans_up_after_synthesized_finish_reason_chunk():
 
 
 @pytest.mark.asyncio
-async def test_stream_wrapper_anext_keeps_context_active_through_synthesized_finish_reason_chunk():
+async def test_stream_wrapper_anext_keeps_context_active_through_synthesized_finish_reason_chunk(monkeypatch):
     """Async sibling of test_stream_wrapper_next_keeps_context_active_through_synthesized_finish_reason_chunk -
     _finalize_completed_stream()'s else branch must not restore before
     returning the synthesized chunk either."""
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     trace_id_var.set("outer-trace-anext-finish-reason")
     session_id_var.set("outer-session-anext-finish-reason")
     try:
@@ -4394,6 +4400,7 @@ async def test_stream_wrapper_anext_max_duration_timeout_restores_consumer_corre
     path as every other failure so the consumer's outer correlation context gets
     restored - calling the check before entering __anext__()'s try block would
     let the Timeout bypass that restoration entirely."""
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     monkeypatch.setattr(litellm.constants, "LITELLM_MAX_STREAMING_DURATION_SECONDS", 1)
     trace_id_var.set("outer-trace-max-duration")
     session_id_var.set("outer-session-max-duration")
@@ -4434,12 +4441,13 @@ async def test_stream_wrapper_anext_max_duration_timeout_restores_consumer_corre
 
 
 @pytest.mark.asyncio
-async def test_stream_wrapper_aclose_restores_consumer_correlation_context():
+async def test_stream_wrapper_aclose_restores_consumer_correlation_context(monkeypatch):
     """Explicit early termination (aclose(), e.g. on client disconnect or a
     router fallback aborting an in-progress stream) must restore the caller's
     correlation context too - not just __del__'s best-effort GC-timed fallback,
     since aclose() is normally called deterministically by the consumer/
     framework, unlike __del__."""
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     trace_id_var.set("outer-trace-aclose")
     session_id_var.set("outer-session-aclose")
     try:
@@ -4481,6 +4489,7 @@ async def test_stream_wrapper_aclose_keeps_context_active_through_close_failure_
     branch logs a debug diagnostic. That log line must still carry the
     closing stream's own trace_id/session_id - the outer context must not be
     restored until after the close attempt (and its diagnostic) completes."""
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     trace_id_var.set("outer-trace-close-fail")
     session_id_var.set("outer-session-close-fail")
     try:
@@ -4541,6 +4550,7 @@ def test_handle_stream_fallback_error_restores_context_only_after_exception_mapp
     mapping. The consumer's outer context must not be restored until that
     mapping call returns, or the diagnostic log line would carry the outer
     (or empty) trace_id/session_id instead of the failing stream's own."""
+    monkeypatch.setattr(litellm, "request_correlation_in_logs", True)
     trace_id_var.set("outer-trace-fallback")
     session_id_var.set("outer-session-fallback")
     try:

tests/test_litellm/llms/bedrock/realtime/test_bedrock_realtime_handler.py

@@ -1,3 +1,4 @@
+import asyncio
 import json
 import sys
 import types
@@ -51,6 +52,27 @@ class FakeBedrockStream:
     def __init__(self, input_stream=None):
         self.input_stream = input_stream if input_stream is not None else FakeInputStream()
 
+    async def await_output(self):
+        return (None, EndedBedrockReceiver())
+
+
+class ServiceUnavailableException(Exception):
+    """Named like the modeled AWS SDK error so the handler maps it to HTTP 503"""
+
+
+class ModelStreamErrorException(Exception):
+    """Named like the modeled AWS SDK error so the handler maps it to HTTP 424"""
+
+
+class UnavailableBedrockStream:
+    """Lazy duplex stream whose HTTP response only fails once the output is awaited"""
+
+    def __init__(self):
+        self.input_stream = FakeInputStream()
+
+    async def await_output(self):
+        raise ServiceUnavailableException("fault injected: Bedrock realtime unavailable")
+
 
 class FakeLogging:
     def __init__(self, trace_id="trace-nova-sonic"):
@@ -61,6 +83,7 @@ class DisconnectingClientWS:
     def __init__(self, messages):
         self._messages = list(messages)
         self.sent_to_client = []
+        self.scope = {}
 
     async def receive_text(self):
         if self._messages:
@@ -93,6 +116,7 @@ class RealtimeClientWS:
     def __init__(self):
         self.closed = False
         self.sent_to_client = []
+        self.scope = {}
 
     async def receive_text(self):
         raise RuntimeError("client disconnected")
@@ -104,6 +128,25 @@ class RealtimeClientWS:
         self.closed = True
 
 
+class ConnectedClientWS(RealtimeClientWS):
+    """Client that sends its scripted messages and then stays connected until the server closes it"""
+
+    def __init__(self, messages):
+        super().__init__()
+        self._messages = list(messages)
+        self._closed_event = asyncio.Event()
+
+    async def receive_text(self):
+        if self._messages:
+            return self._messages.pop(0)
+        await self._closed_event.wait()
+        raise RuntimeError("client disconnected")
+
+    async def close(self, code=None, reason=None):
+        self.closed = True
+        self._closed_event.set()
+
+
 class ScriptedBedrockReceiver:
     def __init__(self, payloads):
         self._payloads = list(payloads)
@@ -115,10 +158,48 @@ class ScriptedBedrockReceiver:
         return SimpleNamespace(value=SimpleNamespace(bytes_=payload.encode("utf-8")))
 
 
-class ScriptedBedrockStream:
+class BreakingBedrockReceiver(ScriptedBedrockReceiver):
+    """Delivers its payloads, then the provider stream breaks instead of ending normally"""
+
+    async def receive(self):
+        if not self._payloads:
+            await asyncio.sleep(0)
+            raise ModelStreamErrorException("Nova Sonic stream broke")
+        return await super().receive()
+
+
+class DrainedThenOpenBedrockReceiver(ScriptedBedrockReceiver):
+    """Delivers its payloads, flags `drained`, then stays open like a live Nova Sonic turn"""
+
     def __init__(self, payloads):
+        super().__init__(payloads)
+        self.drained = asyncio.Event()
+
+    async def receive(self):
+        if not self._payloads:
+            self.drained.set()
+            await asyncio.Event().wait()
+        return await super().receive()
+
+
+class ResetOnAudioInputStream(FakeInputStream):
+    """Accepts session setup, then the provider resets the input side once the first response was delivered"""
+
+    def __init__(self, drained):
+        super().__init__()
+        self._drained = drained
+
+    async def send(self, event):
+        if "audioInput" in json.loads(event.value.bytes_.decode("utf-8")).get("event", {}):
+            await self._drained.wait()
+            raise RuntimeError("bedrock input stream reset")
+        self.sent.append(event)
+
+
+class ScriptedBedrockStream:
+    def __init__(self, payloads, receiver_type=ScriptedBedrockReceiver):
         self.input_stream = FakeInputStream()
-        self._receiver = ScriptedBedrockReceiver(payloads)
+        self._receiver = receiver_type(payloads)
 
     async def await_output(self):
         return (None, self._receiver)
@@ -163,6 +244,11 @@ def stub_aws_sdk_client(monkeypatch):
 
         async def invoke_model_with_bidirectional_stream(self, operation_input):
             captured["operation_input"] = operation_input
+            if captured.get("streams"):
+                stream = captured["streams"].pop(0)
+                if isinstance(stream, Exception):
+                    raise stream
+                return stream
             return ScriptedBedrockStream(captured.get("scripted_payloads", []))
 
     package = types.ModuleType("aws_sdk_bedrock_runtime")
@@ -263,7 +349,8 @@ class TestBedrockRealtimeHandler:
             [json.dumps({"type": "session.update", "session": {"instructions": "You are helpful."}})]
         )
 
-        await handler._forward_client_to_bedrock(client_ws, stream, config, "amazon.nova-sonic-v1:0", {})
+        with pytest.raises(RuntimeError, match="bedrock send failed"):
+            await handler._forward_client_to_bedrock(client_ws, stream, config, "amazon.nova-sonic-v1:0", {})
 
         assert stream.input_stream.closed
 
@@ -464,6 +551,154 @@ class TestBedrockRealtimeSessionLifecycle:
         assert client_ws.sent_to_client == []
 
 
+class TestBedrockRealtimeProviderFailurePropagation:
+    """Deferred Nova Sonic failures must escape async_realtime so the router can fall back / cool down (LIT-6484)"""
+
+    SESSION_UPDATE = json.dumps({"type": "session.update", "session": {"instructions": "hi", "modalities": ["text"]}})
+    AWS_PARAMS = {"aws_region_name": "us-east-1", "aws_access_key_id": "k", "aws_secret_access_key": "s"}
+
+    @pytest.mark.asyncio
+    async def test_readiness_failure_escapes_and_fallback_replays_session_update(self, stub_aws_sdk_client):
+        handler = BedrockRealtime()
+        websocket = ConnectedClientWS([self.SESSION_UPDATE])
+        healthy_stream = ScriptedBedrockStream([])
+        eager_failure = ServiceUnavailableException("fault injected before the stream was returned")
+        stub_aws_sdk_client["streams"] = [UnavailableBedrockStream(), eager_failure, healthy_stream]
+
+        with pytest.raises(BedrockError) as failure:
+            await handler.async_realtime(
+                model="amazon.nova-sonic-v1:0", websocket=websocket, logging_obj=FakeLogging(), **self.AWS_PARAMS
+            )
+
+        assert failure.value.status_code == 503
+        assert [json.loads(m)["type"] for m in websocket.sent_to_client] == ["session.created"]
+        assert not websocket.closed, "the proxy route owns the client-facing error event and 1011 close"
+
+        with pytest.raises(ServiceUnavailableException):
+            await handler.async_realtime(
+                model="amazon.nova-sonic-v1:0", websocket=websocket, logging_obj=FakeLogging(), **self.AWS_PARAMS
+            )
+
+        await handler.async_realtime(
+            model="amazon.nova-sonic-v1:0", websocket=websocket, logging_obj=FakeLogging(), **self.AWS_PARAMS
+        )
+
+        assert [json.loads(m)["type"] for m in websocket.sent_to_client] == ["session.created", "session.updated"]
+        replayed = [json.loads(chunk.value.bytes_.decode("utf-8")) for chunk in healthy_stream.input_stream.sent]
+        assert [next(iter(event["event"])) for event in replayed][:2] == ["sessionStart", "promptStart"]
+        assert websocket.closed
+
+    TEXT_TURN = (
+        json.dumps({"event": {"contentStart": {"role": "ASSISTANT", "type": "TEXT"}}}),
+        json.dumps({"event": {"textOutput": {"content": "Hi"}}}),
+        json.dumps({"event": {"contentEnd": {"stopReason": "END_TURN"}}}),
+    )
+
+    @pytest.fixture
+    def spend_dispatch(self, monkeypatch):
+        import litellm.llms.bedrock.realtime.handler as handler_module
+
+        dispatched = {}
+
+        class RecordingLogging(FakeLogging):
+            async def dispatch_success_handlers(self, result=None, prefer_async_handlers=False, **kwargs):
+                dispatched["events"] = result
+
+        class RecordingLoggingWorker:
+            def ensure_initialized_and_enqueue(self, coro):
+                dispatched["coro"] = coro
+
+        monkeypatch.setattr(handler_module, "GLOBAL_LOGGING_WORKER", RecordingLoggingWorker())
+        dispatched["logging_obj"] = RecordingLogging()
+        return dispatched
+
+    @pytest.mark.asyncio
+    async def test_mid_stream_failure_escapes_keeps_partial_spend_and_blocks_replay(
+        self, stub_aws_sdk_client, spend_dispatch
+    ):
+        handler = BedrockRealtime()
+        websocket = ConnectedClientWS([self.SESSION_UPDATE])
+        stream = ScriptedBedrockStream(self.TEXT_TURN, receiver_type=BreakingBedrockReceiver)
+        stub_aws_sdk_client["streams"] = [stream]
+
+        with pytest.raises(BedrockError) as failure:
+            await handler.async_realtime(
+                model="amazon.nova-sonic-v1:0",
+                websocket=websocket,
+                logging_obj=spend_dispatch["logging_obj"],
+                **self.AWS_PARAMS,
+            )
+
+        assert failure.value.status_code == 424
+        await spend_dispatch["coro"]
+        assert [event["type"] for event in spend_dispatch["events"]] == ["response.done"]
+        assert "response.done" in [json.loads(m)["type"] for m in websocket.sent_to_client]
+        flushed = [json.loads(chunk.value.bytes_.decode("utf-8")) for chunk in stream.input_stream.sent]
+        assert [next(iter(event["event"])) for event in flushed][-2:] == ["promptEnd", "sessionEnd"]
+        assert stream.input_stream.closed
+
+        with pytest.raises(BedrockError) as replay:
+            await handler.async_realtime(
+                model="amazon.nova-sonic-v1:0",
+                websocket=websocket,
+                logging_obj=spend_dispatch["logging_obj"],
+                **self.AWS_PARAMS,
+            )
+
+        assert replay.value.status_code == 400, "a committed session must not be silently restarted on a fallback"
+        assert not litellm._should_retry(replay.value.status_code), "the router must not retry the replay refusal"
+        assert "Nova Sonic stream broke" in replay.value.message, "the router surfaces the last attempt's error"
+
+    @pytest.mark.asyncio
+    async def test_input_side_failure_keeps_spend_for_responses_already_delivered(
+        self, stub_aws_sdk_client, spend_dispatch
+    ):
+        receiver = DrainedThenOpenBedrockReceiver(self.TEXT_TURN)
+        stream = ScriptedBedrockStream(self.TEXT_TURN, receiver_type=lambda _payloads: receiver)
+        stream.input_stream = ResetOnAudioInputStream(receiver.drained)
+        stub_aws_sdk_client["streams"] = [stream]
+        websocket = ConnectedClientWS(
+            [self.SESSION_UPDATE, json.dumps({"type": "input_audio_buffer.append", "audio": "AAAA"})]
+        )
+
+        with pytest.raises(RuntimeError, match="bedrock input stream reset"):
+            await BedrockRealtime().async_realtime(
+                model="amazon.nova-sonic-v1:0",
+                websocket=websocket,
+                logging_obj=spend_dispatch["logging_obj"],
+                **self.AWS_PARAMS,
+            )
+
+        assert "response.done" in [json.loads(m)["type"] for m in websocket.sent_to_client]
+        await spend_dispatch["coro"]
+        assert [event["type"] for event in spend_dispatch["events"]] == ["response.done"]
+
+    @pytest.mark.asyncio
+    async def test_stream_failure_after_client_disconnect_is_not_a_provider_failure(self, stub_aws_sdk_client):
+        stream = ScriptedBedrockStream([], receiver_type=BreakingBedrockReceiver)
+        stub_aws_sdk_client["streams"] = [stream]
+
+        await BedrockRealtime().async_realtime(
+            model="amazon.nova-sonic-v1:0", websocket=RealtimeClientWS(), logging_obj=FakeLogging(), **self.AWS_PARAMS
+        )
+
+        assert stream.input_stream.closed
+
+    @pytest.mark.asyncio
+    async def test_session_updated_is_not_sent_before_bedrock_is_ready(self, stub_aws_models):
+        handler = BedrockRealtime()
+        stream = UnavailableBedrockStream()
+        client_ws = DisconnectingClientWS([self.SESSION_UPDATE])
+
+        with pytest.raises(ServiceUnavailableException):
+            await handler._forward_client_to_bedrock(
+                client_ws, stream, BedrockRealtimeConfig(), "amazon.nova-sonic-v1:0", {}, FakeLogging()
+            )
+
+        assert client_ws.sent_to_client == []
+        assert stream.input_stream.closed
+
+
 class TestBedrockRealtimeAwsAuth:
     """AWS auth params passed via litellm_params must reach the Smithy client config (LIT-3923 regression)"""
 

tests/test_litellm/proxy/auth/test_route_checks.py

@@ -1,5 +1,6 @@
 import os
 from datetime import datetime
+from typing import Final
 from unittest.mock import MagicMock, patch
 
 
@@ -3919,10 +3920,15 @@ def test_claude_code_marketplace_routes_open_to_internal_users(route):
 
 
 @pytest.mark.parametrize("user_role", [None, LitellmUserRoles.INTERNAL_USER.value, LitellmUserRoles.INTERNAL_USER_VIEW_ONLY.value])
-def test_auto_router_session_is_reachable_by_any_key_but_benchmarks_stays_admin_only(user_role):
-    valid_token = UserAPIKeyAuth(api_key="hash-of-caller", user_role=user_role)
-    request = MagicMock(spec=Request)
-    request.query_params = {"session_id": "sess-1"}
+@pytest.mark.parametrize("allowed_routes", [None, ["llm_api_routes"]])
+def test_auto_router_session_is_reachable_by_any_key_but_benchmarks_stays_admin_only(
+    user_role: str | None, allowed_routes: list[str] | None
+) -> None:
+    valid_token: Final = UserAPIKeyAuth(api_key="hash-of-caller", user_role=user_role, allowed_routes=allowed_routes)
+    request: Final = Request({"type": "http", "method": "GET", "query_string": b"session_id=sess-1"})
+
+    assert RouteChecks.should_call_route("/auto_router/session", valid_token, request) is True
+    assert RouteChecks.is_llm_api_route("/auto_router/session") is False
 
     RouteChecks.non_proxy_admin_allowed_routes_check(
         user_obj=None,
@@ -3941,3 +3947,36 @@ def test_auto_router_session_is_reachable_by_any_key_but_benchmarks_stays_admin_
             valid_token=valid_token,
             request_data={},
         )
+
+
+@pytest.mark.parametrize(
+    "route,method,allowed_routes",
+    [
+        ("/auto_router/session", method, ["llm_api_routes"])
+        for method in ("POST", "PUT", "PATCH", "DELETE", "HEAD", "OPTIONS", None)
+    ]
+    + [
+        (route, "GET", ["llm_api_routes"])
+        for route in (
+            "/auto_router/benchmarks",
+            "/auto_router/test_routing",
+            "/auto_router/validate_complexity_router_config",
+            "/auto_router/session/other",
+            "/auto_router/sessions",
+        )
+    ]
+    + [
+        ("/auto_router/session", "GET", allowed_routes)
+        for allowed_routes in (["/v1/messages"], ["info_routes"], ["openai_routes"])
+    ],
+)
+def test_auto_router_session_read_grant_rejects_other_methods_paths_and_scopes(
+    route: str, method: str | None, allowed_routes: list[str]
+) -> None:
+    valid_token: Final = UserAPIKeyAuth(api_key="hash-of-caller", allowed_routes=allowed_routes)
+    request: Final = Request({"type": "http", "method": method}) if method is not None else None
+
+    with pytest.raises(HTTPException) as error:
+        RouteChecks.should_call_route(route, valid_token, request)
+
+    assert error.value.status_code == 403

tests/test_litellm/proxy/auth/test_user_api_key_auth.py

@@ -13,7 +13,7 @@ from unittest.mock import ANY, AsyncMock, MagicMock, patch
 
 
 import pytest
-from fastapi import status
+from fastapi import HTTPException, status
 
 import litellm
 import litellm.proxy.proxy_server
@@ -7628,3 +7628,226 @@ async def test_claude_view_never_reinterprets_explicit_names(monkeypatch, layer)
     assert data["model"] == ("foo" if layer == "unclaimed" else encoded)
     await _normalize_claude_model(data, token, request, "/v1/messages")
     assert data["model"] == ("foo" if layer == "unclaimed" else encoded)
+
+
+ISSUER_ONE = "https://issuer-one.example.com"
+ISSUER_TWO = "https://issuer-two.example.com"
+
+
+def _per_issuer_virtual_key_jwt_handler(
+    global_claim_field: str | None, global_behavior: str = "fallback_team_mapping"
+) -> MagicMock:
+    jwt_handler = MagicMock()
+    jwt_handler.is_jwt.return_value = True
+    jwt_handler.litellm_jwtauth = LiteLLM_JWTAuth(
+        virtual_key_claim_field=global_claim_field,
+        unregistered_jwt_client_behavior=global_behavior,
+        issuers=[
+            {
+                "issuer": ISSUER_ONE,
+                "jwks_url": f"{ISSUER_ONE}/keys",
+                "audience": "audience-one",
+                "team_id_jwt_field": "sub",
+            },
+            {
+                "issuer": ISSUER_TWO,
+                "jwks_url": f"{ISSUER_TWO}/keys",
+                "audience": "audience-two",
+                "virtual_key_claim_field": "sub",
+                "unregistered_jwt_client_behavior": "reject",
+            },
+        ],
+    )
+    return jwt_handler
+
+
+def _fake_prisma_with_jwt_key_mapping(hashed_token: str | None) -> tuple[SimpleNamespace, AsyncMock]:
+    find_first = AsyncMock(return_value=None if hashed_token is None else SimpleNamespace(token=hashed_token))
+    prisma_client = SimpleNamespace(db=SimpleNamespace(litellm_jwtkeymapping=SimpleNamespace(find_first=find_first)))
+    return prisma_client, find_first
+
+
+def _mapping_where(claim_name: str, claim_value: str) -> dict[str, str | bool]:
+    return {"jwt_claim_name": claim_name, "jwt_claim_value": claim_value, "is_active": True}
+
+
+@pytest.mark.asyncio
+async def test_per_issuer_virtual_key_claim_field_selects_the_issuer_mapping_for_the_db_lookup():
+    from litellm.proxy.auth.user_api_key_auth import _resolve_jwt_to_virtual_key
+
+    jwt_handler = _per_issuer_virtual_key_jwt_handler(global_claim_field=None)
+    prisma_client, find_first = _fake_prisma_with_jwt_key_mapping("hashed-mapped-key")
+    user_api_key_cache = DualCache()
+    await user_api_key_cache.async_set_cache(
+        key="hashed-mapped-key",
+        value=UserAPIKeyAuth(token="hashed-mapped-key", api_key="hashed-mapped-key", team_id="svc-team"),
+    )
+
+    resolved = await _resolve_jwt_to_virtual_key(
+        jwt_claims={JWTHandler.LITELLM_JWT_ISSUER_CLAIM: ISSUER_TWO, "sub": "svc-account-7"},
+        jwt_handler=jwt_handler,
+        prisma_client=prisma_client,
+        user_api_key_cache=user_api_key_cache,
+        parent_otel_span=None,
+        proxy_logging_obj=MagicMock(),
+    )
+
+    find_first.assert_awaited_once_with(where=_mapping_where("sub", "svc-account-7"))
+    assert isinstance(resolved, UserAPIKeyAuth)
+    assert resolved.token == "hashed-mapped-key"
+    assert resolved.team_id == "svc-team"
+    assert await user_api_key_cache.async_get_cache("jwt_key_mapping:sub:svc-account-7") == "hashed-mapped-key"
+
+
+@pytest.mark.asyncio
+async def test_per_issuer_reject_behavior_does_not_leak_into_the_team_issuer():
+    from litellm.proxy.auth.user_api_key_auth import _resolve_jwt_to_virtual_key
+
+    jwt_handler = _per_issuer_virtual_key_jwt_handler(global_claim_field=None)
+    prisma_client, find_first = _fake_prisma_with_jwt_key_mapping(None)
+
+    team_issuer_result = await _resolve_jwt_to_virtual_key(
+        jwt_claims={JWTHandler.LITELLM_JWT_ISSUER_CLAIM: ISSUER_ONE, "sub": "team-alpha"},
+        jwt_handler=jwt_handler,
+        prisma_client=prisma_client,
+        user_api_key_cache=DualCache(),
+        parent_otel_span=None,
+        proxy_logging_obj=MagicMock(),
+    )
+    assert team_issuer_result is None
+    find_first.assert_not_awaited()
+
+    with pytest.raises(HTTPException) as exc:
+        await _resolve_jwt_to_virtual_key(
+            jwt_claims={JWTHandler.LITELLM_JWT_ISSUER_CLAIM: ISSUER_TWO, "sub": "unknown-svc"},
+            jwt_handler=jwt_handler,
+            prisma_client=prisma_client,
+            user_api_key_cache=DualCache(),
+            parent_otel_span=None,
+            proxy_logging_obj=MagicMock(),
+        )
+
+    assert exc.value.status_code == 403
+    assert "No registered mapping for sub='unknown-svc'" in str(exc.value.detail)
+    find_first.assert_awaited_once_with(where=_mapping_where("sub", "unknown-svc"))
+
+
+@pytest.mark.asyncio
+async def test_proxy_admin_sentinel_cached_by_another_issuer_does_not_bypass_reject():
+    from litellm.proxy.auth.user_api_key_auth import _JWT_PROXY_ADMIN_SENTINEL, _resolve_jwt_to_virtual_key
+
+    jwt_handler = _per_issuer_virtual_key_jwt_handler(global_claim_field="sub", global_behavior="auto_register")
+    prisma_client, find_first = _fake_prisma_with_jwt_key_mapping(None)
+    user_api_key_cache = DualCache()
+    await user_api_key_cache.async_set_cache(key="jwt_key_mapping:sub:admin-7", value=_JWT_PROXY_ADMIN_SENTINEL)
+
+    auto_register_issuer_result = await _resolve_jwt_to_virtual_key(
+        jwt_claims={JWTHandler.LITELLM_JWT_ISSUER_CLAIM: ISSUER_ONE, "sub": "admin-7"},
+        jwt_handler=jwt_handler,
+        prisma_client=prisma_client,
+        user_api_key_cache=user_api_key_cache,
+        parent_otel_span=None,
+        proxy_logging_obj=MagicMock(),
+    )
+    assert auto_register_issuer_result is None
+    find_first.assert_not_awaited()
+
+    with pytest.raises(HTTPException) as exc:
+        await _resolve_jwt_to_virtual_key(
+            jwt_claims={JWTHandler.LITELLM_JWT_ISSUER_CLAIM: ISSUER_TWO, "sub": "admin-7"},
+            jwt_handler=jwt_handler,
+            prisma_client=prisma_client,
+            user_api_key_cache=user_api_key_cache,
+            parent_otel_span=None,
+            proxy_logging_obj=MagicMock(),
+        )
+
+    assert exc.value.status_code == 403
+    assert "No registered mapping for sub='admin-7'" in str(exc.value.detail)
+    find_first.assert_awaited_once_with(where=_mapping_where("sub", "admin-7"))
+
+
+@pytest.mark.asyncio
+async def test_issuer_without_virtual_key_claim_field_falls_back_to_the_global_field():
+    from litellm.proxy.auth.user_api_key_auth import _resolve_jwt_to_virtual_key
+
+    jwt_handler = _per_issuer_virtual_key_jwt_handler(global_claim_field="client_id")
+    prisma_client, find_first = _fake_prisma_with_jwt_key_mapping(None)
+
+    with_claim = await _resolve_jwt_to_virtual_key(
+        jwt_claims={JWTHandler.LITELLM_JWT_ISSUER_CLAIM: ISSUER_ONE, "sub": "team-alpha", "client_id": "app-9"},
+        jwt_handler=jwt_handler,
+        prisma_client=prisma_client,
+        user_api_key_cache=DualCache(),
+        parent_otel_span=None,
+        proxy_logging_obj=MagicMock(),
+    )
+    without_claim = await _resolve_jwt_to_virtual_key(
+        jwt_claims={JWTHandler.LITELLM_JWT_ISSUER_CLAIM: ISSUER_ONE, "sub": "team-alpha"},
+        jwt_handler=jwt_handler,
+        prisma_client=prisma_client,
+        user_api_key_cache=DualCache(),
+        parent_otel_span=None,
+        proxy_logging_obj=MagicMock(),
+    )
+
+    assert with_claim is None
+    assert without_claim is None
+    find_first.assert_awaited_once_with(where=_mapping_where("client_id", "app-9"))
+
+
+@pytest.mark.asyncio
+async def test_auth_flow_enters_virtual_key_mapping_when_only_an_issuer_configures_the_claim_field():
+    jwt_token = "eyJhbGciOiJSUzI1NiJ9.eyJzdWIiOiJzdmMtYWNjb3VudC03In0.signature"
+    jwt_handler = _per_issuer_virtual_key_jwt_handler(global_claim_field=None)
+    jwt_handler.auth_jwt = AsyncMock(
+        return_value={JWTHandler.LITELLM_JWT_ISSUER_CLAIM: ISSUER_TWO, "sub": "svc-account-7"}
+    )
+    mapped_key = UserAPIKeyAuth(token="hashed-mapped-key", api_key="hashed-mapped-key", team_id="svc-team")
+
+    mock_request = MagicMock()
+    mock_request.url.path = "/v1/chat/completions"
+    mock_request.method = "POST"
+    mock_request.headers = {"authorization": f"Bearer {jwt_token}"}
+    mock_request.query_params = {}
+    mock_request.state = SimpleNamespace()
+
+    with (
+        patch(  # test-quality-ok: the builder reads proxy settings from module globals, no injection seam
+            "litellm.proxy.proxy_server.general_settings", {"enable_jwt_auth": True}
+        ),
+        patch("litellm.proxy.proxy_server.premium_user", True),  # test-quality-ok: module-global proxy state
+        patch("litellm.proxy.proxy_server.master_key", "sk-master"),  # test-quality-ok: module-global proxy state
+        patch("litellm.proxy.proxy_server.prisma_client", MagicMock()),  # test-quality-ok: module-global proxy state
+        patch(  # test-quality-ok: module-global proxy state
+            "litellm.proxy.proxy_server.user_api_key_cache", DualCache()
+        ),
+        patch(  # test-quality-ok: module-global proxy state
+            "litellm.proxy.proxy_server.proxy_logging_obj", MagicMock()
+        ),
+        patch("litellm.proxy.proxy_server.jwt_handler", jwt_handler),  # test-quality-ok: module-global proxy state
+        patch(  # test-quality-ok: the regression is whether the builder reaches this seam at all
+            "litellm.proxy.auth.user_api_key_auth._resolve_jwt_to_virtual_key",
+            new_callable=AsyncMock,
+            return_value=mapped_key,
+        ) as resolve_mock,
+        patch(  # test-quality-ok: a mapped key must short-circuit standard JWT auth; reaching it is the failure
+            "litellm.proxy.auth.user_api_key_auth.JWTAuthManager.auth_builder",
+            new_callable=AsyncMock,
+            side_effect=AssertionError("standard JWT auth must not run for a mapped virtual key"),
+        ),
+    ):
+        result = await _user_api_key_auth_builder(
+            request=mock_request,
+            api_key=jwt_token,
+            azure_api_key_header="",
+            anthropic_api_key_header=None,
+            google_ai_studio_api_key_header=None,
+            azure_apim_header=None,
+            request_data={"model": "gpt-4o-mini"},
+        )
+
+    resolve_mock.assert_awaited_once()
+    assert resolve_mock.await_args.kwargs["jwt_claims"][JWTHandler.LITELLM_JWT_ISSUER_CLAIM] == ISSUER_TWO
+    assert result.api_key == "hashed-mapped-key"
+    assert result.team_id == "svc-team"

tests/test_litellm/proxy/client/cli/test_statusline_script.py

@@ -8,6 +8,7 @@ import re
 import subprocess
 import sys
 from pathlib import Path
+from typing import Final
 
 import pytest
 
@@ -297,16 +298,31 @@ class TestRender:
 
 
 class TestClaudeCodeMode:
-    def test_the_transcript_names_the_routed_model_and_the_proxy_adds_the_savings(self, tmp_path, transcript, config_dir):
-        seen = []
+    @pytest.mark.parametrize("transcript_model", ("claude-auto", "anthropic/claude-opus-5"))
+    def test_the_session_names_the_routed_model_even_when_the_transcript_differs(
+        self, tmp_path: Path, config_dir: Path, transcript_model: str
+    ) -> None:
+        transcript: Final = tmp_path / "session.jsonl"
+        transcript.write_text(_assistant_line(transcript_model) + "\n")
 
-        def fetch(credentials, session_id):
-            seen.append((credentials, session_id))
+        def fetch(credentials: Credentials, session_id: str) -> Fetched:
+            assert credentials == Credentials("http://127.0.0.1:4000", "sk-virtual")
+            assert session_id == SESSION_ID
             return Fetched(RECORDED, definitive=True)
 
-        text = _run(_payload(transcript), _env(tmp_path, config_dir), fetch)
+        text: Final = _run(_payload(transcript), _env(tmp_path, config_dir), fetch)
         assert text.startswith("claude-auto · Routed to: claude-sonnet-5  -63% vs Claude Opus 5\n")
-        assert seen == [(Credentials("http://127.0.0.1:4000", "sk-virtual"), SESSION_ID)]
+
+    def test_a_discovered_display_name_labels_the_sessions_model(
+        self, tmp_path: Path, transcript: Path, config_dir: Path
+    ) -> None:
+        session: Final = RECORDED._replace(last_model="anthropic/claude-opus-5")
+
+        def fetch(credentials: Credentials, session_id: str) -> Fetched:
+            return Fetched(session, definitive=True)
+
+        text: Final = _run(_payload(transcript), _env(tmp_path, config_dir), fetch)
+        assert text.startswith("claude-auto · Routed to: Claude Opus 5  -63% vs Claude Opus 5\n")
 
     def test_an_unrecorded_session_degrades_to_the_routed_line(self, tmp_path, transcript, config_dir):
         assert _run(_payload(transcript), _env(tmp_path, config_dir), lambda c, s: Fetched(None, True)) == (

tests/test_litellm/proxy/discovery_endpoints/test_ui_discovery_endpoints.py

@@ -340,6 +340,7 @@ def test_ui_discovery_endpoints_hide_default_credentials_hint_default_false():
         patch.dict(os.environ, {"DISABLE_ADMIN_UI": "false"}, clear=False),
     ):
         os.environ.pop("LITELLM_HIDE_DEFAULT_CREDENTIALS_HINT", None)
+        os.environ.pop("UI_PASSWORD", None)
 
         response = client.get("/.well-known/litellm-ui-config")
 
@@ -348,6 +349,43 @@ def test_ui_discovery_endpoints_hide_default_credentials_hint_default_false():
         assert data["hide_default_credentials_hint"] is False
 
 
+def test_ui_discovery_endpoints_hide_default_credentials_hint_when_ui_password_set():
+    app = FastAPI()
+    app.include_router(router)
+    client = TestClient(app)
+
+    with patch.dict(os.environ, {"UI_PASSWORD": "s3cret-pass", "DISABLE_ADMIN_UI": "false"}, clear=False):
+        os.environ.pop("LITELLM_HIDE_DEFAULT_CREDENTIALS_HINT", None)
+
+        response = client.get("/.well-known/litellm-ui-config")
+
+        assert response.status_code == 200
+        assert response.json()["hide_default_credentials_hint"] is True
+
+
+@pytest.mark.parametrize(
+    "env_overrides",
+    [
+        pytest.param({"UI_USERNAME": "opsadmin"}, id="username_only_keeps_master_key_password"),
+        pytest.param({"UI_PASSWORD": ""}, id="empty_password_is_not_set"),
+    ],
+)
+def test_ui_discovery_endpoints_keeps_default_credentials_hint_without_real_ui_password(env_overrides):
+    app = FastAPI()
+    app.include_router(router)
+    client = TestClient(app)
+
+    with patch.dict(os.environ, {"DISABLE_ADMIN_UI": "false", **env_overrides}, clear=False):
+        os.environ.pop("LITELLM_HIDE_DEFAULT_CREDENTIALS_HINT", None)
+        if "UI_PASSWORD" not in env_overrides:
+            os.environ.pop("UI_PASSWORD", None)
+
+        response = client.get("/.well-known/litellm-ui-config")
+
+        assert response.status_code == 200
+        assert response.json()["hide_default_credentials_hint"] is False
+
+
 def test_ui_discovery_endpoints_hide_default_credentials_hint_via_env_var():
     """LITELLM_HIDE_DEFAULT_CREDENTIALS_HINT=true hides the login-page credentials card."""
     app = FastAPI()

tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py

@@ -3937,6 +3937,7 @@ async def test_list_team_v2_org_admin_sees_org_teams():
         mock_db.litellm_teamtable.find_many = AsyncMock(return_value=[mock_team])
         mock_db.litellm_teamtable.count = AsyncMock(return_value=1)
         mock_db.litellm_verificationtoken.group_by = AsyncMock(return_value=[])
+        mock_db.litellm_usertable.find_unique = AsyncMock(return_value=mock_user)
 
         result = await list_team_v2(
             http_request=mock_request,
@@ -4036,6 +4037,7 @@ async def test_list_team_v2_org_admin_own_user_id_sees_all_org_teams():
         )
         mock_db.litellm_teamtable.count = AsyncMock(return_value=2)
         mock_db.litellm_verificationtoken.group_by = AsyncMock(return_value=[])
+        mock_db.litellm_usertable.find_unique = AsyncMock(return_value=mock_user)
 
         # UI sends the caller's own user_id for non-Admin roles
         result = await list_team_v2(
@@ -4055,10 +4057,217 @@ async def test_list_team_v2_org_admin_own_user_id_sees_all_org_teams():
         assert result["total"] == 2
         assert len(result["teams"]) == 2
 
-        # Verify the where clause scopes by org only — no team_id filter
+        # Verify the where clause scopes by org OR own membership — no
+        # top-level team_id filter that would hide org teams they aren't in
         where = mock_db.litellm_teamtable.find_many.call_args.kwargs["where"]
-        assert where["organization_id"] == {"in": ["org_A"]}
+        assert where["AND"] == [
+            {"OR": [{"organization_id": {"in": ["org_A"]}}, {"team_id": {"in": ["team_1"]}}]}
+        ]
         assert "team_id" not in where
+        assert "organization_id" not in where
+
+
+def _team_where_matches(team, where) -> bool:
+    for key, cond in where.items():
+        if key == "AND":
+            if not all(_team_where_matches(team, c) for c in cond):
+                return False
+        elif key == "OR":
+            if not any(_team_where_matches(team, c) for c in cond):
+                return False
+        else:
+            value = getattr(team, key)
+            if not isinstance(cond, dict):
+                if value != cond:
+                    return False
+            elif "in" in cond and value not in cond["in"]:
+                return False
+            elif "contains" in cond and cond["contains"].lower() not in (value or "").lower():
+                return False
+    return True
+
+
+def _org_membership(user_id: str, organization_id: str, user_role: str) -> LiteLLM_OrganizationMembershipTable:
+    return LiteLLM_OrganizationMembershipTable(
+        user_id=user_id,
+        organization_id=organization_id,
+        user_role=user_role,
+        spend=0.0,
+        created_at=datetime.now(),
+        updated_at=datetime.now(),
+    )
+
+
+@pytest.mark.asyncio
+async def test_list_team_v2_org_admin_own_query_keeps_memberships_in_other_orgs(monkeypatch):
+    """
+    /v2/team/list: an org admin of org_A who is a member of a team in org_B
+    gets that team back on a self query (with and without user_id, with and
+    without search), alongside every org_A team. The membership half of the
+    union comes from the DB, so a stale cached user object cannot hide it.
+    A query for another user stays scoped to org_A.
+
+    Regression test for LIT-3723.
+    """
+    from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache
+    from litellm.proxy.management_endpoints.team_endpoints import list_team_v2
+
+    org_admin = UserAPIKeyAuth(user_role=LitellmUserRoles.INTERNAL_USER, user_id="org_admin_user")
+    cache = UserApiKeyCache()
+    await cache.async_set_cache(
+        key="org_admin_user",
+        value=LiteLLM_UserTable(
+            user_id="org_admin_user",
+            teams=["team_in_org_A"],
+            organization_memberships=[
+                _org_membership("org_admin_user", "org_A", "org_admin"),
+                _org_membership("org_admin_user", "org_B", "internal_user"),
+            ],
+        ),
+        model_type=LiteLLM_UserTable,
+    )
+    await cache.async_set_cache(
+        key="other_user",
+        value=LiteLLM_UserTable(
+            user_id="other_user",
+            teams=["other_team_in_org_A", "team_in_org_B", "unrelated_team_in_org_B"],
+            organization_memberships=[_org_membership("other_user", "org_B", "internal_user")],
+        ),
+        model_type=LiteLLM_UserTable,
+    )
+
+    def team(team_id, organization_id, *member_ids):
+        return LiteLLM_TeamTable(
+            team_id=team_id,
+            team_alias=team_id,
+            organization_id=organization_id,
+            members_with_roles=[Member(user_id=m, role="user") for m in member_ids],
+        )
+
+    all_teams = [
+        team("team_in_org_A", "org_A", "org_admin_user"),
+        team("other_team_in_org_A", "org_A", "other_user"),
+        team("team_in_org_B", "org_B", "org_admin_user", "other_user"),
+        team("unrelated_team_in_org_B", "org_B", "other_user"),
+    ]
+
+    async def find_many(where=None, **kwargs):
+        return [t for t in all_teams if where is None or _team_where_matches(t, where)]
+
+    async def count(where=None, **kwargs):
+        return len(await find_many(where))
+
+    prisma_client = MagicMock()
+    prisma_client.db.litellm_teamtable.find_many = AsyncMock(side_effect=find_many)
+    prisma_client.db.litellm_teamtable.count = AsyncMock(side_effect=count)
+    prisma_client.db.litellm_verificationtoken.group_by = AsyncMock(return_value=[])
+    prisma_client.db.litellm_usertable.find_unique = AsyncMock(
+        return_value=LiteLLM_UserTable(
+            user_id="org_admin_user",
+            teams=["team_in_org_A", "team_in_org_B"],
+            organization_memberships=[
+                _org_membership("org_admin_user", "org_A", "org_admin"),
+                _org_membership("org_admin_user", "org_B", "internal_user"),
+            ],
+        )
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", prisma_client)
+    monkeypatch.setattr("litellm.proxy.proxy_server.user_api_key_cache", cache)
+    proxy_logging_obj = MagicMock()
+    proxy_logging_obj.service_logging_obj.async_service_success_hook = AsyncMock()
+    monkeypatch.setattr("litellm.proxy.proxy_server.proxy_logging_obj", proxy_logging_obj)
+
+    async def list_teams(user_id, search=None):
+        result = await list_team_v2(
+            http_request=MagicMock(),
+            user_id=user_id,
+            organization_id=None,
+            team_id=None,
+            team_alias=None,
+            search=search,
+            user_api_key_dict=org_admin,
+            page=1,
+            page_size=10,
+            sort_by=None,
+            sort_order="asc",
+            status=None,
+        )
+        assert result["total"] == len(result["teams"])
+        return [t.team_id for t in result["teams"]]
+
+    own_view = ["team_in_org_A", "other_team_in_org_A", "team_in_org_B"]
+    assert await list_teams("org_admin_user") == own_view
+    assert await list_teams(None) == own_view
+    assert await list_teams("org_admin_user", search="team_in_org_B") == ["team_in_org_B"]
+    assert await list_teams("other_user") == ["other_team_in_org_A"]
+    prisma_client.db.litellm_usertable.find_unique.assert_awaited_with(
+        where={"user_id": "org_admin_user"}, include={"organization_memberships": True}
+    )
+
+    prisma_client.db.litellm_usertable.find_unique.side_effect = RuntimeError("db down")
+    with pytest.raises(ValueError, match="db down"):
+        await list_teams("org_admin_user")
+
+
+@pytest.mark.asyncio
+async def test_list_team_v1_org_admin_own_query_keeps_memberships_in_other_orgs():
+    """
+    /team/list: an org admin of org_A listing their own teams sees every team
+    they belong to, including the org_B one. The bare admin listing stays the
+    org_A view and a query for another user stays scoped to org_A.
+
+    Regression test for LIT-3723.
+    """
+    from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache
+    from litellm.proxy.management_endpoints.team_endpoints import _authorize_and_filter_teams
+
+    org_admin = UserAPIKeyAuth(user_role=LitellmUserRoles.INTERNAL_USER, user_id="org_admin_user")
+    cache = UserApiKeyCache()
+    await cache.async_set_cache(
+        key="org_admin_user",
+        value=LiteLLM_UserTable(
+            user_id="org_admin_user",
+            teams=["team_in_org_A", "team_in_org_B"],
+            organization_memberships=[_org_membership("org_admin_user", "org_A", "org_admin")],
+        ),
+        model_type=LiteLLM_UserTable,
+    )
+
+    def team(team_id, organization_id, *member_ids):
+        return SimpleNamespace(
+            team_id=team_id,
+            organization_id=organization_id,
+            members_with_roles=[{"user_id": m, "role": "user"} for m in member_ids],
+        )
+
+    all_teams = [
+        team("team_in_org_A", "org_A", "org_admin_user"),
+        team("other_team_in_org_A", "org_A", "other_user"),
+        team("team_in_org_B", "org_B", "org_admin_user", "other_user"),
+        team("unrelated_team_in_org_B", "org_B", "other_user"),
+    ]
+
+    async def find_many(where=None, **kwargs):
+        if where is None:
+            return all_teams
+        return [t for t in all_teams if t.organization_id in where["organization_id"]["in"]]
+
+    prisma_client = MagicMock()
+    prisma_client.db.litellm_teamtable.find_many = AsyncMock(side_effect=find_many)
+
+    async def list_teams(user_id):
+        teams = await _authorize_and_filter_teams(
+            user_api_key_dict=org_admin,
+            user_id=user_id,
+            prisma_client=prisma_client,
+            user_api_key_cache=cache,
+            proxy_logging_obj=MagicMock(),
+        )
+        return [t.team_id for t in teams]
+
+    assert await list_teams("org_admin_user") == ["team_in_org_A", "team_in_org_B"]
+    assert await list_teams(None) == ["team_in_org_A", "other_team_in_org_A"]
+    assert await list_teams("other_user") == ["other_team_in_org_A"]
 
 
 @pytest.mark.asyncio

tests/test_litellm/proxy/management_endpoints/test_ui_sso.py

@@ -8334,6 +8334,7 @@ async def _render_legacy_login_page(env_overrides, general_settings):
             "GOOGLE_CLIENT_ID",
             "GENERIC_CLIENT_ID",
             "LITELLM_HIDE_DEFAULT_CREDENTIALS_HINT",
+            "UI_PASSWORD",
         ):
             os.environ.pop(var, None)
         os.environ.update(env_overrides)
@@ -8386,6 +8387,20 @@ async def test_legacy_login_page_hides_credentials_hint_via_general_settings():
     assert "MASTER_KEY" not in body
 
 
+@pytest.mark.asyncio
+async def test_legacy_login_page_hides_credentials_hint_when_ui_password_set():
+    response = await _render_legacy_login_page(
+        env_overrides={"UI_PASSWORD": "s3cret-pass"},
+        general_settings={},
+    )
+
+    body = response.body.decode()
+    assert response.status_code == 200
+    assert "Default Credentials" not in body
+    assert "MASTER_KEY" not in body
+    assert 'name="username"' in body
+
+
 @pytest.mark.asyncio
 async def test_saml_callback_blocked_when_admin_ui_disabled():
     """An IdP-initiated assertion must not mint a UI session when the admin UI is

tests/test_litellm/proxy/openai_files_endpoint/test_files_endpoint.py

@@ -4703,6 +4703,108 @@ def test_create_file_blocked_extension_unset_allows_everything(monkeypatch, llm_
     assert len(forwarded_calls) == 1
 
 
+@pytest.mark.parametrize(
+    "filename",
+    ["payload.exe", "notes.txt", "README"],
+    ids=["other_extension", "text_extension", "no_extension"],
+)
+def test_create_file_extension_outside_allowlist_rejected_before_forwarding(
+    monkeypatch, llm_router: Router, filename: str
+):
+    import litellm.proxy.proxy_server as ps
+
+    forwarded_calls = _setup_batch_upload_endpoint(monkeypatch, llm_router)
+    monkeypatch.setitem(ps.general_settings, "allowed_file_extensions", [".jsonl"])
+
+    try:
+        response = client.post(
+            "/v1/files",
+            files={"file": (filename, b"MZ\x90\x00", "application/octet-stream")},
+            data={"purpose": "user_data"},
+            headers={"Authorization": "Bearer test-key"},
+        )
+    finally:
+        _teardown_batch_upload_endpoint()
+
+    assert response.status_code == 400, response.text
+    error = response.json()["error"]
+    assert error["type"] == "invalid_request_error"
+    assert error["param"] == "file"
+    assert "allowed_file_extensions" in error["message"]
+    assert forwarded_calls == []
+
+
+def test_create_file_allowed_extension_forwards_case_insensitively(monkeypatch, llm_router: Router):
+    import litellm.proxy.proxy_server as ps
+
+    forwarded_calls = _setup_batch_upload_endpoint(monkeypatch, llm_router)
+    monkeypatch.setitem(ps.general_settings, "allowed_file_extensions", [".JSONL"])
+
+    try:
+        response = client.post(
+            "/v1/files",
+            files={"file": ("input.jsonl", b'{"custom_id": "1"}\n', "application/jsonl")},
+            data={"purpose": "user_data"},
+            headers={"Authorization": "Bearer test-key"},
+        )
+    finally:
+        _teardown_batch_upload_endpoint()
+
+    assert response.status_code == 200, response.text
+    assert len(forwarded_calls) == 1
+
+
+def test_create_file_empty_allowlist_rejects_every_upload(monkeypatch, llm_router: Router):
+    import litellm.proxy.proxy_server as ps
+
+    forwarded_calls = _setup_batch_upload_endpoint(monkeypatch, llm_router)
+    monkeypatch.setitem(ps.general_settings, "allowed_file_extensions", [])
+
+    try:
+        response = client.post(
+            "/v1/files",
+            files={"file": ("input.jsonl", b'{"custom_id": "1"}\n', "application/jsonl")},
+            data={"purpose": "user_data"},
+            headers={"Authorization": "Bearer test-key"},
+        )
+    finally:
+        _teardown_batch_upload_endpoint()
+
+    assert response.status_code == 400, response.text
+    assert "allowed_file_extensions" in response.json()["error"]["message"]
+    assert forwarded_calls == []
+
+
+def test_create_file_allowlist_runs_before_blocklist(monkeypatch, llm_router: Router):
+    import litellm.proxy.proxy_server as ps
+
+    forwarded_calls = _setup_batch_upload_endpoint(monkeypatch, llm_router)
+    monkeypatch.setitem(ps.general_settings, "allowed_file_extensions", [".jsonl"])
+    monkeypatch.setitem(ps.general_settings, "blocked_file_extensions", [".exe", ".jsonl"])
+
+    try:
+        denied_by_allowlist = client.post(
+            "/v1/files",
+            files={"file": ("payload.exe", b"MZ\x90\x00", "application/octet-stream")},
+            data={"purpose": "user_data"},
+            headers={"Authorization": "Bearer test-key"},
+        )
+        denied_by_blocklist = client.post(
+            "/v1/files",
+            files={"file": ("input.jsonl", b'{"custom_id": "1"}\n', "application/jsonl")},
+            data={"purpose": "user_data"},
+            headers={"Authorization": "Bearer test-key"},
+        )
+    finally:
+        _teardown_batch_upload_endpoint()
+
+    assert denied_by_allowlist.status_code == 400, denied_by_allowlist.text
+    assert "allowed_file_extensions" in denied_by_allowlist.json()["error"]["message"]
+    assert denied_by_blocklist.status_code == 400, denied_by_blocklist.text
+    assert "blocked_file_extensions" in denied_by_blocklist.json()["error"]["message"]
+    assert forwarded_calls == []
+
+
 def test_create_file_path_traversal_filename_rejected_before_forwarding(monkeypatch, llm_router: Router):
     """A filename carrying a directory-traversal component must never reach storage or the provider."""
     forwarded_calls = _setup_batch_upload_endpoint(monkeypatch, llm_router)

tests/test_litellm/proxy/openai_files_endpoint/test_general_upload_validation.py

@@ -1,4 +1,5 @@
 import io
+from pathlib import Path
 
 import pytest
 
@@ -6,11 +7,14 @@ from litellm.proxy._types import ProxyException
 from litellm.proxy.openai_files_endpoints.general_upload_validation import (
     MB,
     UploadedFileBlockedExtension,
+    UploadedFileExtensionNotAllowed,
     UploadedFileTooLarge,
     UploadedFileUnsafeFilename,
+    check_allowed_extension,
     check_blocked_extension,
     check_unsafe_filename,
     check_upload_file_size,
+    coerce_optional_str_list_setting,
     raise_upload_validation_failure,
 )
 
@@ -79,6 +83,44 @@ def test_no_filename_skips_extension_check():
     assert check_blocked_extension(None, (".exe",)) is None
 
 
+def test_allowed_extension_passes():
+    assert check_allowed_extension("batch.jsonl", (".jsonl", ".pdf")) is None
+
+
+@pytest.mark.parametrize("filename", ["payload.exe", "notes.txt", "archive.tar.gz"])
+def test_extension_outside_allowlist_rejected(filename):
+    assert check_allowed_extension(filename, (".jsonl", ".pdf")) == UploadedFileExtensionNotAllowed(
+        extension=Path(filename).suffix
+    )
+
+
+def test_allowed_extension_match_is_case_insensitive_for_upload():
+    assert check_allowed_extension("batch.JSONL", (".jsonl",)) is None
+
+
+def test_allowed_extension_match_is_case_insensitive_for_configured_value():
+    assert check_allowed_extension("batch.jsonl", (".JSONL",)) is None
+
+
+@pytest.mark.parametrize("filename", ["README", "", None, "../../"])
+def test_no_extension_rejected_when_allowlist_set(filename):
+    assert check_allowed_extension(filename, (".jsonl",)) == UploadedFileExtensionNotAllowed(extension="")
+
+
+def test_empty_allowlist_rejects_everything():
+    assert check_allowed_extension("batch.jsonl", ()) == UploadedFileExtensionNotAllowed(extension=".jsonl")
+
+
+def test_unset_allowlist_skips_check():
+    assert check_allowed_extension("payload.exe", None) is None
+
+
+def test_coerce_str_list_setting_keeps_unset_and_empty_distinct():
+    assert coerce_optional_str_list_setting(None) is None
+    assert coerce_optional_str_list_setting([]) == ()
+    assert coerce_optional_str_list_setting([".jsonl"]) == (".jsonl",)
+
+
 def test_path_traversal_filename_rejected():
     assert check_unsafe_filename("../../etc/passwd") == UploadedFileUnsafeFilename(filename="../../etc/passwd")
 
@@ -112,6 +154,16 @@ def test_ordinary_filenames_allowed(filename):
             "413",
             ("15.0 MB", "max_file_size_mb", "10 MB", "not forwarded"),
         ),
+        (
+            UploadedFileExtensionNotAllowed(extension=".exe"),
+            "400",
+            (".exe", "allowed_file_extensions", "not forwarded"),
+        ),
+        (
+            UploadedFileExtensionNotAllowed(extension=""),
+            "400",
+            ("without an extension", "allowed_file_extensions", "not forwarded"),
+        ),
         (
             UploadedFileBlockedExtension(extension=".exe"),
             "400",

tests/test_litellm/proxy/proxy_server/test_proxy_config.py

@@ -3469,6 +3469,30 @@ async def test_ProxyConfig__update_general_settings_cleared_db_max_batch_file_si
     assert ps.general_settings.get("max_batch_file_size_mb") is None
 
 
+@pytest.mark.asyncio
+async def test_ProxyConfig__update_general_settings_applies_db_allowed_file_extensions(monkeypatch):
+    monkeypatch.setattr("litellm.proxy.proxy_server.general_settings", {})
+    pc = ProxyConfig()
+    await pc._update_general_settings({"allowed_file_extensions": [".jsonl"]})
+    from litellm.proxy import proxy_server as ps
+
+    assert ps.general_settings.get("allowed_file_extensions") == [".jsonl"]
+
+
+@pytest.mark.asyncio
+async def test_ProxyConfig__update_general_settings_yaml_allowed_file_extensions_wins_over_db(monkeypatch):
+    monkeypatch.setattr(
+        "litellm.proxy.proxy_server.general_settings",
+        {"allowed_file_extensions": [".pdf"]},
+    )
+    pc = ProxyConfig()
+    pc._yaml_general_settings_keys = {"allowed_file_extensions"}
+    await pc._update_general_settings({"allowed_file_extensions": [".jsonl"]})
+    from litellm.proxy import proxy_server as ps
+
+    assert ps.general_settings.get("allowed_file_extensions") == [".pdf"]
+
+
 @pytest.mark.asyncio
 async def test_ProxyConfig__update_general_settings_none_input_noop():
     pc = ProxyConfig()

tests/test_litellm/proxy/proxy_server/test_routes_login_sso.py

@@ -97,6 +97,7 @@ def test_fallback_login_returns_html_form_with_ui_username_set(client, monkeypat
 def test_fallback_login_shows_credentials_hint_by_default(client, monkeypatch):
     """Control: without the flag, /fallback/login still renders the hint."""
     monkeypatch.delenv("UI_USERNAME", raising=False)
+    monkeypatch.delenv("UI_PASSWORD", raising=False)
     monkeypatch.delenv("LITELLM_HIDE_DEFAULT_CREDENTIALS_HINT", raising=False)
     response = client.get("/fallback/login")
     assert response.status_code == 200
@@ -104,6 +105,16 @@ def test_fallback_login_shows_credentials_hint_by_default(client, monkeypatch):
     assert "MASTER_KEY" in response.text
 
 
+def test_fallback_login_hides_credentials_hint_when_ui_password_set(client, monkeypatch):
+    monkeypatch.setenv("UI_PASSWORD", "s3cret-pass")
+    monkeypatch.delenv("LITELLM_HIDE_DEFAULT_CREDENTIALS_HINT", raising=False)
+    response = client.get("/fallback/login")
+    assert response.status_code == 200
+    assert "Default Credentials" not in response.text
+    assert "MASTER_KEY" not in response.text
+    assert 'name="username"' in response.text
+
+
 def test_fallback_login_hides_credentials_hint_via_env_flag(client, monkeypatch):
     """Pin: LITELLM_HIDE_DEFAULT_CREDENTIALS_HINT removes the hint on /fallback/login."""
     monkeypatch.delenv("UI_USERNAME", raising=False)

tests/test_litellm/proxy/proxy_server/test_spend_counters.py

@@ -1180,6 +1180,24 @@ async def test_apply_spend_counter_increments_open_breaker_invalidates_and_retur
     fake_cache.in_memory_cache.set_cache.assert_not_called()
 
 
+@pytest.mark.asyncio
+async def test_apply_spend_counter_increments_redis_timeout_invalidates_and_returns(monkeypatch):
+    """A Redis timeout invalidates the counters and returns without reaching the cost callback's error path."""
+    from redis.exceptions import TimeoutError as RedisTimeoutError
+
+    fake_cache = _make_spend_counter_cache()
+    fake_cache.redis_cache.async_increment_pipeline = AsyncMock(
+        side_effect=RedisTimeoutError("Timeout reading from 127.0.0.1:6379")
+    )
+    monkeypatch.setattr(ps, "spend_counter_cache", fake_cache)
+
+    await ps._apply_spend_counter_increments(_two_pending_increments())
+
+    deleted_keys = sorted(call.kwargs["key"] for call in fake_cache.in_memory_cache.delete_cache.call_args_list)
+    assert deleted_keys == ["spend:key:k", "spend:team:t"]
+    fake_cache.in_memory_cache.set_cache.assert_not_called()
+
+
 @pytest.mark.asyncio
 async def test_apply_spend_counter_increments_other_redis_error_invalidates_and_raises(monkeypatch):
     fake_cache = _make_spend_counter_cache()

tests/test_litellm/proxy/proxy_server/test_team_model_name_translation.py

@@ -9,6 +9,7 @@ rows instead of the internal routing key `model_name_{team_id}_{uuid}`.
 
 from __future__ import annotations
 
+import json
 from unittest.mock import AsyncMock, MagicMock
 
 import pytest
@@ -238,7 +239,7 @@ async def test_model_info_v1_list_path_translates_team_model_name(monkeypatch):
     )
     resp = await ps.model_info_v1(user_api_key_dict=admin, litellm_model_id=None)
 
-    names = [m["model_name"] for m in resp["data"]]
+    names = [m["model_name"] for m in json.loads(resp.body)["data"]]
     assert "team-claude-sonnet" in names
     assert "model_name_team-abc-123_4a6b8" not in names
 
@@ -271,7 +272,7 @@ async def test_model_info_v1_unrestricted_key_returns_all_deployments(monkeypatc
     )
     resp = await ps.model_info_v1(user_api_key_dict=caller, litellm_model_id=None)
 
-    assert [m["model_name"] for m in resp["data"]] == ["gpt-4"]
+    assert [m["model_name"] for m in json.loads(resp.body)["data"]] == ["gpt-4"]
 
 
 @pytest.mark.asyncio
@@ -303,7 +304,7 @@ async def test_model_info_v1_restricted_key_filters_deployments(monkeypatch):
     )
     resp = await ps.model_info_v1(user_api_key_dict=caller, litellm_model_id=None)
 
-    assert [m["model_name"] for m in resp["data"]] == ["gpt-4"]
+    assert [m["model_name"] for m in json.loads(resp.body)["data"]] == ["gpt-4"]
 
 
 def _other_team_row() -> dict:
@@ -367,10 +368,11 @@ async def test_model_info_v1_unrestricted_key_hides_other_team_byok(monkeypatch)
     )
     resp = await ps.model_info_v1(user_api_key_dict=caller, litellm_model_id=None)
 
-    returned_ids = {m["model_info"]["id"] for m in resp["data"]}
+    data = json.loads(resp.body)["data"]
+    returned_ids = {m["model_info"]["id"] for m in data}
     assert returned_ids == {"global-id-1", "byok-id-1"}
     assert "byok-id-other" not in returned_ids
-    names = [m["model_name"] for m in resp["data"]]
+    names = [m["model_name"] for m in data]
     assert "team-claude-sonnet" in names
     assert "gpt-4" in names
 
@@ -412,7 +414,7 @@ async def test_model_info_v1_service_key_hides_all_team_byok(monkeypatch):
     )
     resp = await ps.model_info_v1(user_api_key_dict=caller, litellm_model_id=None)
 
-    assert [m["model_info"]["id"] for m in resp["data"]] == ["global-id-1"]
+    assert [m["model_info"]["id"] for m in json.loads(resp.body)["data"]] == ["global-id-1"]
 
 
 @pytest.mark.asyncio
@@ -466,7 +468,7 @@ async def test_model_info_v1_team_key_sees_own_byok_regardless_of_user_lookup(
     )
     resp = await ps.model_info_v1(user_api_key_dict=caller, litellm_model_id=None)
 
-    assert [m["model_info"]["id"] for m in resp["data"]] == ["byok-id-1", "global-id-1"]
+    assert [m["model_info"]["id"] for m in json.loads(resp.body)["data"]] == ["byok-id-1", "global-id-1"]
 
 
 @pytest.mark.asyncio
@@ -509,7 +511,7 @@ async def test_model_info_v1_user_team_membership_grants_byok(monkeypatch):
     )
     resp = await ps.model_info_v1(user_api_key_dict=caller, litellm_model_id=None)
 
-    assert [m["model_info"]["id"] for m in resp["data"]] == [
+    assert [m["model_info"]["id"] for m in json.loads(resp.body)["data"]] == [
         "byok-id-other",
         "global-id-1",
     ]
@@ -557,7 +559,7 @@ async def test_model_info_v1_populates_access_via_team_ids(monkeypatch):
     )
     resp = await ps.model_info_v1(user_api_key_dict=admin, litellm_model_id=None)
 
-    by_id = {m["model_info"]["id"]: m for m in resp["data"]}
+    by_id = {m["model_info"]["id"]: m for m in json.loads(resp.body)["data"]}
     assert by_id["byok-id-1"]["model_info"]["access_via_team_ids"] == [team_id]
     assert by_id["byok-id-1"]["model_info"]["direct_access"] is False
     assert by_id["global-id-1"]["model_info"]["direct_access"] is True
@@ -816,7 +818,7 @@ async def test_model_info_v1_litellm_model_id_include_team_models_filters_inacce
         include_team_models=True,
     )
 
-    assert resp["data"] == []
+    assert json.loads(resp.body)["data"] == []
 
 
 @pytest.mark.asyncio
@@ -852,7 +854,7 @@ async def test_model_info_v1_litellm_model_id_team_id_applies_team_filter(monkey
         teamId="other-team",
     )
 
-    assert resp["data"] == []
+    assert json.loads(resp.body)["data"] == []
     team_filter.assert_awaited_once()
     assert team_filter.await_args.kwargs["team_id"] == "other-team"
     assert team_filter.await_args.kwargs["all_models"] == [team_row]

tests/test_litellm/proxy/test__types.py

@@ -277,3 +277,61 @@ def test_team_membership_budget_table_present_still_works():
     }
     result = LiteLLM_TeamMembership.model_validate(data)
     assert result.litellm_budget_table is None
+
+
+def test_a_jwt_issuer_can_override_the_virtual_key_claim_field_while_other_issuers_keep_the_global_one():
+    from litellm.proxy._types import LiteLLM_JWTAuth, UnregisteredJWTClientBehavior
+
+    jwt_auth = LiteLLM_JWTAuth(
+        virtual_key_claim_field="client_id",
+        issuers=[
+            {
+                "issuer": "https://team-idp.example.com",
+                "jwks_url": "https://team-idp.example.com/keys",
+                "audience": "litellm",
+                "team_id_jwt_field": "sub",
+            },
+            {
+                "issuer": "https://service-idp.example.com",
+                "jwks_url": "https://service-idp.example.com/keys",
+                "audience": "litellm",
+                "virtual_key_claim_field": "sub",
+                "unregistered_jwt_client_behavior": "reject",
+            },
+        ],
+    )
+
+    assert jwt_auth.get_virtual_key_claim_field("https://service-idp.example.com") == "sub"
+    assert jwt_auth.get_unregistered_jwt_client_behavior("https://service-idp.example.com") is (
+        UnregisteredJWTClientBehavior.REJECT
+    )
+    assert jwt_auth.get_virtual_key_claim_field("https://team-idp.example.com") == "client_id"
+    assert jwt_auth.get_unregistered_jwt_client_behavior("https://team-idp.example.com") is (
+        UnregisteredJWTClientBehavior.FALLBACK_TEAM_MAPPING
+    )
+    assert jwt_auth.get_virtual_key_claim_field(None) == "client_id"
+    assert jwt_auth.get_virtual_key_claim_field("https://unknown-idp.example.com") == "client_id"
+
+
+@pytest.mark.parametrize(
+    ("global_field", "issuer_field", "is_configured"),
+    ((None, None, False), ("sub", None, True), (None, "sub", True)),
+)
+def test_virtual_key_mapping_counts_as_configured_when_any_issuer_sets_the_claim_field(
+    global_field, issuer_field, is_configured
+):
+    from litellm.proxy._types import LiteLLM_JWTAuth
+
+    jwt_auth = LiteLLM_JWTAuth(
+        virtual_key_claim_field=global_field,
+        issuers=[
+            {
+                "issuer": "https://idp.example.com",
+                "jwks_url": "https://idp.example.com/keys",
+                "audience": "litellm",
+                "virtual_key_claim_field": issuer_field,
+            }
+        ],
+    )
+
+    assert jwt_auth.is_virtual_key_mapping_configured() is is_configured

tests/test_litellm/proxy/test_common_request_processing.py

@@ -3879,6 +3879,39 @@ class TestHandleLLMApiExceptionRetryAfter:
         assert proxy_exc.headers["retry-after"] == "43"
         assert proxy_exc.headers["x-custom"] == "1"
 
+    async def test_handle_llm_api_exception_names_cooldown_when_every_deployment_is_cooled_down(self):
+        from litellm.types.router import RouterRateLimitError
+
+        exc = RouterRateLimitError(
+            model="gpt-4",
+            cooldown_time=120,
+            enable_pre_call_checks=False,
+            cooldown_list=["dep-a", "dep-b"],
+            model_ids=["dep-a", "dep-b"],
+        )
+        proxy_exc = await self._invoke(exc)
+        body = proxy_exc.to_dict()
+        assert body["type"] == "all_deployments_in_cooldown"
+        assert body["code"] == "429"
+        assert "All deployments for selected model are in cooldown" in body["message"]
+        assert proxy_exc.headers["retry-after"] == "120"
+
+    async def test_handle_llm_api_exception_keeps_rate_limit_type_when_cooldown_is_partial(self):
+        from litellm.types.router import RouterRateLimitError
+
+        exc = RouterRateLimitError(
+            model="gpt-4",
+            cooldown_time=120,
+            enable_pre_call_checks=False,
+            cooldown_list=["dep-a"],
+            model_ids=["dep-a", "dep-b"],
+        )
+        proxy_exc = await self._invoke(exc)
+        body = proxy_exc.to_dict()
+        assert body["type"] == "rate_limit_error"
+        assert body["code"] == "429"
+        assert "All deployments for selected model are in cooldown" not in body["message"]
+
 
 class TestHandleLLMApiExceptionFramingHeaders:
     """HTTP-framing headers on the provider exception must be stripped before the
@@ -8400,6 +8433,41 @@ async def test_handle_llm_api_exception_forwards_provider_headers_on_http_status
     assert exc_info.value.headers["llm_provider-x-amzn-requestid"] == "req-passthrough-500"
 
 
+@pytest.mark.asyncio
+async def test_handle_llm_api_exception_forwards_litellm_response_headers_when_response_is_synthetic():
+    """Exception mapping hands the proxy a mapped error whose ``response`` is a synthetic empty
+    ``httpx.Response`` and parks the provider's real headers on ``litellm_response_headers``.
+    The client must still get the provider request id, as it does on a 200.
+    """
+    import httpx
+
+    from litellm.proxy._types import ProxyException, UserAPIKeyAuth
+
+    mapped = litellm.BadRequestError(
+        message="OpenAIException - max_tokens is too large: 999999999.",
+        model="gpt-4o-mini",
+        llm_provider="openai",
+    )
+    mapped.litellm_response_headers = httpx.Headers({"x-request-id": "req_openai_400"})
+    assert dict(mapped.response.headers) == {}
+
+    processor = ProxyBaseLLMRequestProcessing(data={})
+    proxy_logging_obj = MagicMock()
+    proxy_logging_obj.post_call_failure_hook = AsyncMock(return_value=None)
+    proxy_logging_obj.post_call_response_headers_hook = AsyncMock(return_value={})
+
+    with pytest.raises(ProxyException) as exc_info:
+        await processor._handle_llm_api_exception(
+            e=mapped,
+            user_api_key_dict=UserAPIKeyAuth(api_key="sk-test"),
+            proxy_logging_obj=proxy_logging_obj,
+        )
+
+    assert exc_info.value.code == "400"
+    assert "max_tokens is too large: 999999999." in exc_info.value.message
+    assert exc_info.value.headers["llm_provider-x-request-id"] == "req_openai_400"
+
+
 class TestBackgroundResponseRetrievalGovernance:
     """LIT-7175: retrieving a background Response attaches the model's post_call policy pipelines."""
 

tests/test_litellm/proxy/test_model_info_default_limits.py

@@ -3,6 +3,7 @@ Tests verifying that default_api_key_tpm_limit and default_api_key_rpm_limit set
 litellm_params are returned by the /model/info endpoint.
 """
 
+import json
 from typing import Optional
 from unittest.mock import MagicMock, patch
 
@@ -128,8 +129,9 @@ class TestModelInfoEndpointWithRouter:
                 litellm_model_id="some-model-id",
             )
 
-        assert len(response["data"]) == 1
-        litellm_params = response["data"][0]["litellm_params"]
+        data = json.loads(response.body)["data"]
+        assert len(data) == 1
+        litellm_params = data[0]["litellm_params"]
         assert litellm_params.get("default_api_key_tpm_limit") == 100
         assert litellm_params.get("default_api_key_rpm_limit") == 200
 
@@ -171,7 +173,8 @@ class TestModelInfoEndpointWithRouter:
                 litellm_model_id=None,
             )
 
-        assert len(response["data"]) >= 1
-        litellm_params = response["data"][0]["litellm_params"]
+        data = json.loads(response.body)["data"]
+        assert len(data) >= 1
+        litellm_params = data[0]["litellm_params"]
         assert litellm_params.get("default_api_key_tpm_limit") == 100
         assert litellm_params.get("default_api_key_rpm_limit") == 200

tests/test_litellm/proxy/test_model_list_healthy_only.py

@@ -6,6 +6,7 @@ per-request `healthy_only` query parameter and the proxy-wide
 (`model_info_v1`).
 """
 
+import json
 from unittest.mock import AsyncMock, MagicMock
 
 import pytest
@@ -275,7 +276,7 @@ async def test_model_info_v1_healthy_only_hides_unhealthy_deployments(
         litellm_model_id=None,
         healthy_only=True,
     )
-    assert [m["model_name"] for m in response["data"]] == ["gpt-4"]
+    assert [m["model_name"] for m in json.loads(response.body)["data"]] == ["gpt-4"]
 
 
 @pytest.mark.asyncio
@@ -286,7 +287,7 @@ async def test_model_info_v1_general_setting_hides_unhealthy_deployments(patched
         user_api_key_dict=_admin_key(),
         litellm_model_id=None,
     )
-    assert [m["model_name"] for m in response["data"]] == ["gpt-4"]
+    assert [m["model_name"] for m in json.loads(response.body)["data"]] == ["gpt-4"]
 
 
 @pytest.mark.asyncio
@@ -297,7 +298,7 @@ async def test_model_info_v1_default_keeps_unhealthy_deployments(
         user_api_key_dict=_admin_key(),
         litellm_model_id=None,
     )
-    assert [m["model_name"] for m in response["data"]] == ["gpt-4", "claude-sonnet"]
+    assert [m["model_name"] for m in json.loads(response.body)["data"]] == ["gpt-4", "claude-sonnet"]
     patched_model_info_v1.async_get_fully_unhealthy_model_names.assert_not_awaited()
 
 
@@ -318,4 +319,4 @@ async def test_model_info_v1_litellm_model_id_lookup_ignores_health_filter(patch
         user_api_key_dict=_admin_key(),
         litellm_model_id="unhealthy-id",
     )
-    assert [m["model_name"] for m in response["data"]] == ["claude-sonnet"]
+    assert [m["model_name"] for m in json.loads(response.body)["data"]] == ["claude-sonnet"]

tests/test_litellm/proxy/test_proxy_server.py

@@ -15,10 +15,12 @@ from unittest import mock
 from unittest.mock import AsyncMock, MagicMock, create_autospec, mock_open, patch
 
 import click
+import fastapi.routing
 import httpx
 import pytest
 import yaml
 from fastapi import FastAPI
+from fastapi.encoders import jsonable_encoder
 from fastapi.staticfiles import StaticFiles
 from fastapi.testclient import TestClient
 
@@ -5247,6 +5249,8 @@ async def test_model_info_v1_oci_secrets_not_leaked():
         result = await model_info_v1(user_api_key_dict=mock_user_api_key_dict, litellm_model_id=None)
 
         # Verify the result structure
+        result_str = result.body.decode()
+        result = json.loads(result_str)
         assert "data" in result
         assert len(result["data"]) == 1
 
@@ -5269,13 +5273,96 @@ async def test_model_info_v1_oci_secrets_not_leaked():
         assert litellm_params["model"].startswith("oci/"), "model should retain its full value"
 
         # Verify that actual secret values are not present in the response
-        result_str = str(result)
         assert "ocid1.api_key.oc1..aaaaaaaa7kbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbk" not in result_str
         assert "aa:bb:cc:dd:ee:ff:11:22:33:44:55:66:77:88:99:00" not in result_str
         assert "ocid1.tenancy.oc1..aaaaaaaa7kbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbkbk" not in result_str
         assert "/path/to/oci_api_key.pem" not in result_str
 
 
+def test_model_info_v1_list_skips_fastapi_jsonable_encoder(monkeypatch):
+    """
+    /model/info serializes its multi-megabyte listing itself with orjson. FastAPI must not
+    re-walk the payload through `jsonable_encoder`, while values orjson cannot encode natively
+    still come out as JSON.
+    """
+    created_at = datetime(2026, 1, 2, 3, 4, 5, tzinfo=timezone.utc)
+    model_data = {
+        "model_name": "gpt-4o",
+        "litellm_params": {"model": "openai/gpt-4o", "api_key": "sk-secret-value"},
+        "model_info": {
+            "id": "db-row-1",
+            "db_model": True,
+            "created_at": created_at,
+            "supported_regions": frozenset({"eu"}),
+        },
+    }
+    mock_router = MagicMock()
+    mock_router.model_list = [model_data]
+    mock_router.get_model_list_from_model_alias.return_value = []
+    mock_router.get_model_names.return_value = ["gpt-4o"]
+    mock_router.get_model_access_groups.return_value = {}
+    mock_router.get_deployment.return_value = None
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", mock_router)
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_model_list", [model_data])
+    monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", None)
+    monkeypatch.setattr("litellm.proxy.proxy_server.general_settings", {"infer_model_from_keys": False})
+    monkeypatch.setattr("litellm.proxy.proxy_server.user_model", None)
+
+    encoder_spy = MagicMock(wraps=jsonable_encoder)
+    monkeypatch.setattr(fastapi.routing, "jsonable_encoder", encoder_spy)
+
+    original_overrides = app.dependency_overrides.copy()
+    app.dependency_overrides[user_api_key_auth] = lambda: UserAPIKeyAuth(
+        user_role=LitellmUserRoles.PROXY_ADMIN, api_key="sk-1234", models=[], team_models=[]
+    )
+    client = TestClient(app)
+    try:
+        response = client.get("/model/info")
+    finally:
+        app.dependency_overrides = original_overrides
+
+    assert response.status_code == 200
+    assert response.headers["content-type"] == "application/json"
+    rows = response.json()["data"]
+    assert [row["model_name"] for row in rows] == ["gpt-4o"]
+    assert rows[0]["model_info"]["created_at"] == created_at.isoformat()
+    assert rows[0]["model_info"]["supported_regions"] == ["eu"]
+    assert "sk-secret-value" not in response.text
+    assert encoder_spy.call_count == 0
+
+
+def test_model_info_v1_cli_model_returns_single_deployment_as_json(monkeypatch):
+    """
+    A proxy started with `litellm --model <name>` answers /model/info with one deployment
+    object under `data`, serialized the same way as the listing.
+    """
+    monkeypatch.setattr("litellm.proxy.proxy_server.user_model", "gpt-4o")
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None)
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_model_list", None)
+    monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", None)
+    monkeypatch.setattr("litellm.proxy.proxy_server.general_settings", {})
+
+    encoder_spy = MagicMock(wraps=jsonable_encoder)
+    monkeypatch.setattr(fastapi.routing, "jsonable_encoder", encoder_spy)
+
+    original_overrides = app.dependency_overrides.copy()
+    app.dependency_overrides[user_api_key_auth] = lambda: UserAPIKeyAuth(
+        user_role=LitellmUserRoles.PROXY_ADMIN, api_key="sk-1234", models=[], team_models=[]
+    )
+    client = TestClient(app)
+    try:
+        response = client.get("/model/info")
+    finally:
+        app.dependency_overrides = original_overrides
+
+    assert response.status_code == 200
+    assert response.headers["content-type"] == "application/json"
+    deployment = response.json()["data"]
+    assert deployment["model_name"] == "*"
+    assert deployment["litellm_params"]["model"] == "gpt-4o"
+    assert encoder_spy.call_count == 0
+
+
 def test_add_callback_from_db_to_in_memory_litellm_callbacks():
     """
     Test that _add_callback_from_db_to_in_memory_litellm_callbacks correctly adds callbacks

tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py

@@ -2789,6 +2789,7 @@ class TestUsageTransformation:
         assert response_usage.input_tokens_details is not None
         assert response_usage.input_tokens_details.cached_tokens == 5
         assert response_usage.input_tokens_details.text_tokens == 8
+        assert "cache_write_tokens" not in response_usage.input_tokens_details.model_dump()
 
     def test_transform_usage_with_cached_tokens_gemini(self):
         """Test that cached_tokens from Gemini are properly transformed to input_tokens_details"""
@@ -2851,6 +2852,7 @@ class TestUsageTransformation:
         assert response_usage.input_tokens_details is not None
         assert response_usage.input_tokens_details.cached_tokens == 100
         assert getattr(response_usage.input_tokens_details, "cache_write_tokens", None) == 800
+        assert response_usage.input_tokens_details.model_dump()["cache_write_tokens"] == 800
 
     def test_transform_usage_with_reasoning_tokens_gemini(self):
         """Test that reasoning_tokens from Gemini are properly transformed to output_tokens_details"""

tests/test_litellm/responses/test_responses_utils.py

@@ -577,6 +577,47 @@ class TestResponseAPILoggingUtils:
         assert result.completion_tokens_details is not None
         assert result.completion_tokens_details.reasoning_tokens == 4
 
+    def test_transform_realtime_usage_dict_keeps_cached_tokens_details(self):
+        usage = {
+            "input_tokens": 283,
+            "output_tokens": 0,
+            "total_tokens": 283,
+            "input_token_details": {
+                "text_tokens": 116,
+                "audio_tokens": 167,
+                "cached_tokens": 192,
+                "cached_tokens_details": {"text_tokens": 64, "audio_tokens": 128},
+            },
+        }
+
+        result = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(usage)
+
+        assert result.prompt_tokens_details is not None
+        assert result.prompt_tokens_details.cached_tokens == 192
+        assert result.prompt_tokens_details.cached_tokens_details is not None
+        assert result.prompt_tokens_details.cached_tokens_details.audio_tokens == 128
+        assert result.prompt_tokens_details.cached_tokens_details.text_tokens == 64
+
+    def test_transform_response_api_usage_object_keeps_cached_tokens_details(self):
+        usage = ResponseAPIUsage(
+            input_tokens=283,
+            output_tokens=0,
+            total_tokens=283,
+            input_tokens_details={
+                "text_tokens": 116,
+                "audio_tokens": 167,
+                "cached_tokens": 192,
+                "cached_tokens_details": {"text_tokens": 64, "audio_tokens": 128},
+            },
+        )
+
+        result = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(usage)
+
+        assert result.prompt_tokens_details is not None
+        assert result.prompt_tokens_details.cached_tokens_details is not None
+        assert result.prompt_tokens_details.cached_tokens_details.audio_tokens == 128
+        assert result.prompt_tokens_details.cached_tokens_details.text_tokens == 64
+
 
 class TestResponsesAPIProviderSpecificParams:
     """

tests/test_litellm/rust_bridge/test_lifecycle.py

@@ -0,0 +1,30 @@
+from typing import Final
+
+import pytest
+
+import litellm
+from litellm.rust_bridge.lifecycle import check_limits
+
+
+@pytest.mark.parametrize("metadata_key", ["metadata", "litellm_metadata"])
+@pytest.mark.parametrize(
+    "cap, attempted_retries, refused",
+    [(5, 5, True), (5, 4, False), (0, 0, False), (0, 1, True)],
+    ids=[
+        "cap-above-four-reached",
+        "cap-above-four-not-reached",
+        "first-attempt-passes-cap-of-zero",
+        "cap-of-zero-refuses-first-retry",
+    ],
+)
+def test_check_limits_reads_attempted_retries(
+    monkeypatch: pytest.MonkeyPatch, metadata_key: str, cap: int, attempted_retries: int, refused: bool
+) -> None:
+    monkeypatch.setattr(litellm, "num_retries_per_request", cap)
+    monkeypatch.setattr(litellm, "max_budget", None)
+    kwargs: Final = {"model": "mistral/mistral-ocr-latest", metadata_key: {"attempted_retries": attempted_retries}}
+    if refused:
+        with pytest.raises(RuntimeError, match="Max retries per request hit!"):
+            check_limits(kwargs)
+    else:
+        check_limits(kwargs)

tests/test_litellm/test_cost_calculator.py

@@ -19,6 +19,7 @@ from litellm.cost_calculator import (
 )
 from litellm.litellm_core_utils.litellm_logging import Logging
 from litellm.llms.base_llm.ocr.transformation import OCRPage, OCRResponse, OCRUsageInfo
+from litellm.types.llms.base import CachedTokensDetails
 from litellm.types.llms.openai import OpenAIRealtimeStreamList
 from litellm.types.rerank import RerankResponse
 from litellm.types.utils import (
@@ -4848,6 +4849,109 @@ def test_collect_and_combine_realtime_usage_stores_partitioned_text_tokens() ->
     assert combined.completion_tokens_details.audio_tokens == 0
 
 
+def test_realtime_combine_sums_nested_cached_tokens_details():
+    results: OpenAIRealtimeStreamList = [
+        {
+            "type": "response.done",
+            "response": {
+                "usage": {
+                    "input_tokens": 283,
+                    "output_tokens": 0,
+                    "total_tokens": 283,
+                    "input_token_details": {
+                        "text_tokens": 116,
+                        "audio_tokens": 167,
+                        "cached_tokens": 192,
+                        "cached_tokens_details": {"text_tokens": 64, "audio_tokens": 128},
+                    },
+                }
+            },
+        },
+        {
+            "type": "response.done",
+            "response": {
+                "usage": {
+                    "input_tokens": 150,
+                    "output_tokens": 0,
+                    "total_tokens": 150,
+                    "input_token_details": {
+                        "text_tokens": 50,
+                        "audio_tokens": 100,
+                        "cached_tokens": 100,
+                        "cached_tokens_details": {"audio_tokens": 100},
+                    },
+                }
+            },
+        },
+    ]
+
+    combined = RealtimeAPITokenUsageProcessor.collect_and_combine_usage_from_realtime_stream_results(
+        results=results,
+    )
+
+    assert combined.prompt_tokens_details is not None
+    assert combined.prompt_tokens_details.cached_tokens == 292
+    assert combined.prompt_tokens_details.cached_tokens_details is not None
+    assert combined.prompt_tokens_details.cached_tokens_details.audio_tokens == 228
+    assert combined.prompt_tokens_details.cached_tokens_details.text_tokens == 64
+    assert combined.prompt_tokens_details.cached_tokens_details.image_tokens is None
+
+
+@pytest.mark.parametrize("details_first", [True, False])
+def test_realtime_combine_keeps_cached_split_when_only_one_usage_has_details(details_first: bool):
+    with_details: Final = {
+        "type": "response.done",
+        "response": {
+            "usage": {
+                "input_tokens": 283,
+                "output_tokens": 0,
+                "total_tokens": 283,
+                "input_token_details": {
+                    "text_tokens": 116,
+                    "audio_tokens": 167,
+                    "cached_tokens": 192,
+                    "cached_tokens_details": {"text_tokens": 64, "audio_tokens": 128},
+                },
+            }
+        },
+    }
+    without_details: Final = {
+        "type": "response.done",
+        "response": {
+            "usage": {
+                "input_tokens": 150,
+                "output_tokens": 0,
+                "total_tokens": 150,
+                "input_token_details": {"text_tokens": 50, "audio_tokens": 100, "cached_tokens": 100},
+            }
+        },
+    }
+    results: OpenAIRealtimeStreamList = (
+        [with_details, without_details] if details_first else [without_details, with_details]
+    )
+
+    combined = RealtimeAPITokenUsageProcessor.collect_and_combine_usage_from_realtime_stream_results(
+        results=results,
+    )
+
+    assert combined.prompt_tokens_details is not None
+    assert combined.prompt_tokens_details.cached_tokens == 292
+    assert combined.prompt_tokens_details.cached_tokens_details == CachedTokensDetails(text_tokens=64, audio_tokens=128)
+
+
+def test_usage_without_cached_tokens_details_omits_key():
+    usage = Usage(
+        prompt_tokens=10,
+        completion_tokens=5,
+        total_tokens=15,
+        prompt_tokens_details=PromptTokensDetailsWrapper(text_tokens=10),
+    )
+
+    dumped = usage.prompt_tokens_details.model_dump()
+    assert "cached_tokens_details" not in dumped
+    assert "cached_tokens_details" not in usage.prompt_tokens_details.model_dump_json()
+
+
 UNMAPPED_OCR_MODEL: Final = "azure_ai/some-unmapped-ocr-model-for-testing"
 MAPPED_OCR_MODEL: Final = "mistral/mistral-ocr-4-0"
 

tests/test_litellm/test_logging.py

@@ -17,6 +17,7 @@ from litellm._logging import (
     _MAX_SCRUBBED_ACCESS_ARG,
     _PLAIN_LOG_FORMAT,
     ALL_LOGGERS,
+    AccessLogPathFilter,
     AccessLogRedactionFilter,
     CorrelationContextFilter,
     CorrelationPlainFormatter,
@@ -1178,3 +1179,72 @@ def test_access_redaction_survives_the_uvicorn_json_log_config():
             lg.handlers[:] = handlers
             lg.setLevel(level)
             lg.propagate = True
+
+
+_DISABLED_ACCESS_LOG_PATHS_RAW = " /health/liveliness , ,/metrics/"
+
+
+@pytest.mark.parametrize(
+    "full_path",
+    [
+        "/health/liveliness",
+        "/health/liveliness?x=1",
+        "/health/liveliness?probe=" + "x" * _MAX_SCRUBBED_ACCESS_ARG,
+        "/metrics/",
+        "/metrics/?format=prometheus&job=a",
+    ],
+)
+def test_uvicorn_access_logger_drops_a_configured_path(monkeypatch, full_path):
+    monkeypatch.setenv("LITELLM_DISABLE_ACCESS_LOG_PATHS", _DISABLED_ACCESS_LOG_PATHS_RAW)
+    assert _emit_access_line(full_path) == ""
+
+
+@pytest.mark.parametrize(
+    "full_path",
+    ["/v1/chat/completions", "/health", "/health/liveliness/", "/metrics", "/v1/models?health=/health/liveliness"],
+)
+def test_uvicorn_access_logger_keeps_an_unconfigured_path(monkeypatch, full_path):
+    monkeypatch.setenv("LITELLM_DISABLE_ACCESS_LOG_PATHS", _DISABLED_ACCESS_LOG_PATHS_RAW)
+    assert f'"GET {full_path} HTTP/1.1" 200' in _emit_access_line(full_path)
+
+
+@pytest.mark.parametrize("raw", [None, "", " , ,"])
+def test_uvicorn_access_logger_keeps_every_line_when_no_path_is_configured(monkeypatch, raw):
+    if raw is None:
+        monkeypatch.delenv("LITELLM_DISABLE_ACCESS_LOG_PATHS", raising=False)
+    else:
+        monkeypatch.setenv("LITELLM_DISABLE_ACCESS_LOG_PATHS", raw)
+    assert '"GET /health/liveliness HTTP/1.1" 200' in _emit_access_line("/health/liveliness")
+
+
+def test_access_log_path_filter_survives_the_uvicorn_json_log_config(monkeypatch):
+    import logging.config
+
+    monkeypatch.setenv("LITELLM_DISABLE_ACCESS_LOG_PATHS", _DISABLED_ACCESS_LOG_PATHS_RAW)
+    names = ("uvicorn", "uvicorn.error", "uvicorn.access")
+    saved = tuple((logging.getLogger(n), logging.getLogger(n).handlers[:], logging.getLogger(n).level) for n in names)
+    try:
+        logging.config.dictConfig(_get_uvicorn_json_log_config())
+
+        assert _emit_access_line("/health/liveliness?x=1") == ""
+        assert '"GET /v1/models HTTP/1.1" 200' in _emit_access_line("/v1/models")
+    finally:
+        for lg, handlers, level in saved:
+            lg.handlers[:] = handlers
+            lg.setLevel(level)
+            lg.propagate = True
+
+
+@pytest.mark.parametrize("args", [None, ("127.0.0.1:1", "GET", 42)])
+def test_access_log_path_filter_keeps_a_record_without_a_string_path_arg(monkeypatch, args):
+    monkeypatch.setenv("LITELLM_DISABLE_ACCESS_LOG_PATHS", "/health/liveliness")
+    record = logging.LogRecord(
+        name="uvicorn.access",
+        level=logging.INFO,
+        pathname="",
+        lineno=0,
+        msg='127.0.0.1:1 - "GET /health/liveliness HTTP/1.1" 200',
+        args=args,
+        exc_info=None,
+    )
+    assert AccessLogPathFilter().filter(record) is True

tests/test_litellm/test_model_prices_schema.py

@@ -221,6 +221,14 @@ def test_chat_latest_declares_the_one_effort_openai_accepts(prices: dict):
     assert resolve_supported_reasoning_efforts(prices["chat-latest"], deployment_is_mapped=True) == ("medium",)
 
 
+@pytest.mark.parametrize("key", ["azure/gpt-chat-latest", "azure/chat-latest", "azure/us/gpt-chat-latest"])
+def test_azure_gpt_chat_latest_declares_the_one_effort_azure_accepts(prices: dict, key: str):
+    """Azure answers every reasoning_effort on a gpt-chat-latest deployment except medium with
+    "Unsupported value ... Supported values are: 'medium'", the same fixed level OpenAI's chat-latest
+    carries, so the Foundry product name and the OpenAI API name both declare that one level."""
+    assert resolve_supported_reasoning_efforts(prices[key], deployment_is_mapped=True) == ("medium",)
+
+
 BEDROCK_OPENAI_GPT_MARKERS: Final = ("openai.gpt-5.4", "openai.gpt-5.5", "openai.gpt-5.6", "openai.gpt-6-astra")
 BEDROCK_PROVIDERS: Final = frozenset(("bedrock", "bedrock_converse", "bedrock_mantle"))
 BEDROCK_ROW_PREFIXES: Final = ("bedrock_mantle/", "us.", "global.")

tests/test_litellm/test_router.py

@@ -7735,6 +7735,52 @@ def test_get_available_deployment_raises_when_addressed_dict_is_blocked():
         router.get_available_deployment(model="dep-0", request_kwargs={})
 
 
+def _cool_down(router: Router, *deployment_ids: str) -> None:
+    for deployment_id in deployment_ids:
+        router.cooldown_cache.add_deployment_to_cooldown(
+            model_id=deployment_id,
+            original_exception=litellm.RateLimitError(message="upstream 429", llm_provider="openai", model="gpt-4o"),
+            exception_status=429,
+            cooldown_time=60,
+        )
+
+
+async def _select_deployment(router: Router, use_async: bool) -> None:
+    if use_async:
+        await router.async_get_available_deployment(model="gpt-4o", request_kwargs={})
+        return
+    router.get_available_deployment(model="gpt-4o", request_kwargs={})
+
+
+@pytest.mark.parametrize("use_async", [False, True], ids=["sync", "async"])
+@pytest.mark.asyncio
+async def test_get_available_deployment_names_cooldown_when_every_deployment_is_cooled_down(use_async: bool):
+    from litellm.types.router import RouterErrors, RouterRateLimitError
+
+    router: Final = _router_with_two_deployments([False, False])
+    _cool_down(router, "dep-0", "dep-1")
+    with pytest.raises(RouterRateLimitError) as exc_info:
+        await _select_deployment(router, use_async)
+    assert exc_info.value.all_deployments_in_cooldown is True
+    assert exc_info.value.type == "all_deployments_in_cooldown"
+    assert RouterErrors.all_deployments_in_cooldown.value in str(exc_info.value)
+    assert str(exc_info.value).startswith("No deployments available for selected model, Try again in ")
+
+
+@pytest.mark.parametrize("use_async", [False, True], ids=["sync", "async"])
+@pytest.mark.asyncio
+async def test_get_available_deployment_keeps_generic_error_when_cooldown_is_partial(use_async: bool):
+    from litellm.types.router import RouterErrors, RouterRateLimitError
+
+    router: Final = _router_with_two_deployments([False, True])
+    _cool_down(router, "dep-0")
+    with pytest.raises(RouterRateLimitError) as exc_info:
+        await _select_deployment(router, use_async)
+    assert exc_info.value.all_deployments_in_cooldown is False
+    assert exc_info.value.type == "rate_limit_error"
+    assert RouterErrors.all_deployments_in_cooldown.value not in str(exc_info.value)
+
+
 def _router_with_two_pass_through_deployments(blocked_flags):
     import litellm
 
@@ -7772,6 +7818,24 @@ def test_get_available_deployment_for_pass_through_raises_when_dict_blocked():
         )
 
 
+def test_get_available_deployment_for_pass_through_names_cooldown_despite_healthy_non_pass_through():
+    from litellm.types.router import RouterRateLimitError
+
+    router: Final = _router_with_two_pass_through_deployments([False, False])
+    router.add_deployment(
+        Deployment(
+            model_name="gpt-4o",
+            litellm_params=LiteLLM_Params(model="openai/gpt-4o-plain", api_key="sk-fake-for-tests"),
+            model_info=ModelInfo(id="plain-0"),
+        )
+    )
+    _cool_down(router, "pt-0", "pt-1")
+    with pytest.raises(RouterRateLimitError) as exc_info:
+        router.get_available_deployment_for_pass_through(model="gpt-4o", request_kwargs={})
+    assert exc_info.value.all_deployments_in_cooldown is True
+    assert exc_info.value.type == "all_deployments_in_cooldown"
+
+
 def test_initialize_deployment_for_pass_through_keeps_bedrock_iam_deployment():
     """
     Bedrock deployments using IAM/OIDC auth have no api_key; pass-through
@@ -10642,6 +10706,7 @@ def _cyclic_fallback_router(num_retries=0):
                     "api_key": "sk-fake",
                     "mock_response": "litellm.InternalServerError",
                 },
+                "model_info": {"id": f"{group}-deployment"},
             }
             for group in groups
         ],
@@ -10691,28 +10756,37 @@ async def test_cyclic_fallback_graph_does_not_amplify_one_request():
     assert sum(len(message) for message in capture.messages) < 5_000
 
 
+_FLAT_ATTEMPT_RECORD_KEYS = frozenset(
+    {"model_group", "deployment_id", "exception_type", "exception_string", "attempted_retries"}
+)
+_BREADCRUMB_CREDENTIAL_CANARY = "Bearer sk-ant-oat01-RETRY-BREADCRUMB-CANARY-doNotShip"
+
+
 @pytest.mark.asyncio
-async def test_retry_breadcrumbs_do_not_carry_the_walk_state():
-    """log_retry copies every kwarg into previous_models, which reaches spend logs and
-    logging callbacks. The set of already-attempted groups is router-internal walk state
-    with no diagnostic value there, and it is the one entry that is not a plain scalar.
-    A retry has to be configured for the walk state to reach log_retry at all."""
+async def test_retry_records_are_flat_and_name_the_failed_group_on_fallback_hops():
+    """Each failed attempt leaves a flat record in previous_models, which reaches spend logs and
+    logging callbacks. Nothing downstream reads the failed attempt's kwargs or metadata, and copying
+    them is what carried client credentials and multiplied the payload on every retry. A fallback hop
+    calls log_retry too, so the record has to name the group that failed, not the one taken next."""
     router = _cyclic_fallback_router(num_retries=1)
     capture = _LogCapture(logging.ERROR)
     recorder = _FallbackAttemptRecorder()
 
     await _drive_cyclic_fallback(router, capture, recorder)
 
-    breadcrumbs = [breadcrumb for hop in recorder.breadcrumbs_per_target for breadcrumb in hop]
-    assert breadcrumbs, "no retry breadcrumbs were recorded"
-    assert any(
-        "fallback_depth" in breadcrumb for breadcrumb in breadcrumbs
-    ), "no breadcrumb carried router walk state, so this test cannot see the leak"
-    for breadcrumb in breadcrumbs:
-        assert "attempted_targets" not in breadcrumb
-
-
-_BREADCRUMB_CREDENTIAL_CANARY = "Bearer sk-ant-oat01-RETRY-BREADCRUMB-CANARY-doNotShip"
+    records = [record for hop in recorder.breadcrumbs_per_target for record in hop]
+    assert records, "no retry records were recorded"
+    for record in records:
+        assert set(record) == _FLAT_ATTEMPT_RECORD_KEYS
+        assert record["exception_type"] == "InternalServerError"
+        assert record["deployment_id"] == f"{record['model_group']}-deployment"
+    group_failed_before_hop = {"group-b": "group-a", "group-c": "group-b", "group-d": "group-c"}
+    for failed_target, hop_records in zip(recorder.failed_targets, recorder.breadcrumbs_per_target):
+        groups = [record["model_group"] for record in hop_records]
+        first_own_attempt = groups.index(failed_target)
+        assert groups[first_own_attempt - 1] == group_failed_before_hop[failed_target]
+        assert set(groups[first_own_attempt:]) == {failed_target}
+        assert [record["attempted_retries"] for record in hop_records[first_own_attempt:]][:2] == [0, 1]
 
 
 @pytest.mark.parametrize(
@@ -10738,22 +10812,20 @@ _BREADCRUMB_CREDENTIAL_CANARY = "Bearer sk-ant-oat01-RETRY-BREADCRUMB-CANARY-doN
     ],
 )
 @pytest.mark.asyncio
-async def test_retry_breadcrumbs_never_carry_a_forwarded_credential(container_key, request_kwargs):
-    """log_retry copies kwargs into previous_models, which reaches spend logs and logging callbacks.
-    Any of these kwargs can carry a client's forwarded Authorization token or a provider key, and a
-    breadcrumb has no diagnostic use for the raw secret. A denylist of key names is always one new
-    credential kwarg behind, so log_retry scrubs credential-named values by pattern instead: the
-    container still reaches the breadcrumb, but the raw secret never does, whatever key holds it."""
+async def test_retry_records_never_carry_a_forwarded_credential(container_key, request_kwargs):
+    """previous_models reaches spend logs and logging callbacks. Any request kwarg can carry a client's
+    forwarded Authorization token or a provider key, so the record must not carry request kwargs at
+    all: neither the credential-bearing container nor the raw secret, whatever key holds it."""
     router = _cyclic_fallback_router(num_retries=1)
     capture = _LogCapture(logging.ERROR)
     metadata = {}
 
     await _drive_cyclic_fallback(router, capture, metadata=metadata, **request_kwargs)
 
-    breadcrumbs = metadata["previous_models"]
-    assert breadcrumbs, "no retry breadcrumbs were recorded"
-    dumped = json.dumps(breadcrumbs, default=str)
-    assert container_key in dumped, "the credential-bearing kwarg never reached the breadcrumb, so this test cannot see the leak"
+    records = metadata["previous_models"]
+    assert records, "no retry records were recorded"
+    dumped = json.dumps(records)
+    assert container_key not in dumped
     assert _BREADCRUMB_CREDENTIAL_CANARY not in dumped
 
 
@@ -10773,12 +10845,12 @@ def _always_failing_router(num_retries):
     )
 
 
-async def _fail_one_proxy_shaped_request(router, request_marker):
+async def _fail_one_proxy_shaped_request(router, request_marker, expected_error=litellm.InternalServerError):
     """The proxy hands the router a metadata dict and a proxy_server_request whose body is a
     shallow copy of the request, so body["metadata"] is the very same dict the router later
     stamps previous_models onto."""
     metadata = {"request_marker": request_marker}
-    with pytest.raises(litellm.InternalServerError):
+    with pytest.raises(expected_error):
         await router.acompletion(
             model="broken-group",
             messages=[{"role": "user", "content": "hi"}],
@@ -10804,34 +10876,49 @@ def _nested_breadcrumb_lists(node):
 
 
 @pytest.mark.asyncio
-async def test_retry_breadcrumbs_stay_per_request_and_flat_across_failing_requests():
-    """Every failed attempt appends a breadcrumb to metadata["previous_models"], and the proxy's
+async def test_retry_records_stay_per_request_and_flat_across_failing_requests():
+    """Every failed attempt appends a record to metadata["previous_models"], and the proxy's
     request snapshot aliases that same metadata dict. Kept on the Router and copied wholesale,
-    each breadcrumb embedded every earlier one from every earlier request, so the breadcrumb
+    each breadcrumb once embedded every earlier one from every earlier request, so the breadcrumb
     tree, and with it the debug repr of the kwargs, roughly doubled on each failed attempt until
     a single-worker proxy spent minutes in the redaction regex and stopped answering."""
     router = _always_failing_router(num_retries=2)
 
-    breadcrumbs_per_request = [
+    records_per_request = [
         await _fail_one_proxy_shaped_request(router, f"request-{request_number}") for request_number in range(1, 7)
     ]
 
-    for request_number, breadcrumbs in enumerate(breadcrumbs_per_request, start=1):
-        assert len(breadcrumbs) == 3, "one initial attempt plus two retries failed, each leaving one breadcrumb"
-        assert {breadcrumb["metadata"]["request_marker"] for breadcrumb in breadcrumbs} == {f"request-{request_number}"}
-        for breadcrumb in breadcrumbs:
-            assert _nested_breadcrumb_lists(breadcrumb) == []
-    assert len({len(repr(breadcrumbs)) for breadcrumbs in breadcrumbs_per_request}) == 1
+    for records in records_per_request:
+        assert [record["attempted_retries"] for record in records] == [0, 1, 2]
+        for record in records:
+            assert set(record) == _FLAT_ATTEMPT_RECORD_KEYS
+            assert _nested_breadcrumb_lists(record) == []
+    assert len({len(repr(records)) for records in records_per_request}) == 1
 
 
 @pytest.mark.asyncio
-async def test_retry_breadcrumbs_keep_only_the_last_four_attempts():
+async def test_retry_records_keep_only_the_last_four_attempts():
     router = _always_failing_router(num_retries=6)
 
-    breadcrumbs = await _fail_one_proxy_shaped_request(router, "request-1")
+    records = await _fail_one_proxy_shaped_request(router, "request-1")
 
-    assert len(breadcrumbs) == 4
-    assert [breadcrumb["metadata"]["attempted_retries"] for breadcrumb in breadcrumbs] == [3, 4, 5, 6]
+    assert [record["attempted_retries"] for record in records] == [3, 4, 5, 6]
+
+
+@pytest.mark.asyncio
+async def test_num_retries_per_request_stops_retries_at_caps_above_four(monkeypatch):
+    monkeypatch.setattr(litellm, "num_retries_per_request", 5)
+    router = _always_failing_router(num_retries=6)
+
+    records = await _fail_one_proxy_shaped_request(router, "request-1", expected_error=litellm.APIConnectionError)
+
+    assert [record["attempted_retries"] for record in records] == [3, 4, 5, 6]
+    assert ["Max retries per request hit!" in record["exception_string"] for record in records] == [
+        False,
+        False,
+        True,
+        True,
+    ]
 
 
 @pytest.mark.asyncio

tests/test_litellm/test_utils.py

@@ -7,8 +7,8 @@ import os
 import queue
 import threading
 from datetime import datetime, timedelta, timezone
-from collections.abc import Iterator
-from concurrent.futures import ThreadPoolExecutor
+from collections.abc import Callable, Iterator
+from concurrent.futures import Future, ThreadPoolExecutor
 from typing import Final
 from unittest.mock import AsyncMock, MagicMock, patch
 
@@ -29,6 +29,7 @@ from litellm._logging import (
     verbose_logger,
 )
 from litellm.integrations.custom_logger import CustomLogger
+from litellm.litellm_core_utils.thread_pool_executor import executor as logging_executor
 from litellm.proxy.utils import is_valid_api_key
 from litellm.types.utils import (
     CallTypes,
@@ -4062,6 +4063,51 @@ class TestMetadataNoneHandling:
         assert metadata == {}
 
 
+_RETRY_CAP_CASES: Final = (
+    pytest.param(5, {"attempted_retries": 5}, True, id="cap-above-four-reached"),
+    pytest.param(5, {"attempted_retries": 4}, False, id="cap-above-four-not-reached"),
+    pytest.param(0, {"attempted_retries": 0}, False, id="first-attempt-passes-cap-of-zero"),
+    pytest.param(0, {"attempted_retries": 1}, True, id="cap-of-zero-refuses-first-retry"),
+    pytest.param(5, {"previous_models": ("a", "b", "c", "d", "e")}, False, id="breadcrumb-count-is-not-the-cap"),
+    pytest.param(5, None, False, id="metadata-none"),
+)
+
+
+def _capped_completion_kwargs(metadata_key: str, metadata: object) -> dict[str, object]:
+    return {
+        "model": "openai/gpt-4o-mini",
+        "messages": [{"role": "user", "content": "hi"}],
+        "api_key": "sk-fake",
+        "mock_response": "ok",
+        metadata_key: metadata,
+    }
+
+
+@pytest.mark.parametrize("metadata_key", ["metadata", "litellm_metadata"])
+@pytest.mark.parametrize("cap, metadata, refused", _RETRY_CAP_CASES)
+def test_num_retries_per_request_reads_attempted_retries_sync(monkeypatch, metadata_key, cap, metadata, refused):
+    monkeypatch.setattr(litellm, "num_retries_per_request", cap)
+    kwargs: Final = _capped_completion_kwargs(metadata_key, metadata)
+    if refused:
+        with pytest.raises(Exception, match="Max retries per request hit!"):
+            litellm.completion(**kwargs)
+    else:
+        assert litellm.completion(**kwargs).choices[0].message.content == "ok"
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("metadata_key", ["metadata", "litellm_metadata"])
+@pytest.mark.parametrize("cap, metadata, refused", _RETRY_CAP_CASES)
+async def test_num_retries_per_request_reads_attempted_retries_async(monkeypatch, metadata_key, cap, metadata, refused):
+    monkeypatch.setattr(litellm, "num_retries_per_request", cap)
+    kwargs: Final = _capped_completion_kwargs(metadata_key, metadata)
+    if refused:
+        with pytest.raises(Exception, match="Max retries per request hit!"):
+            await litellm.acompletion(**kwargs)
+    else:
+        assert (await litellm.acompletion(**kwargs)).choices[0].message.content == "ok"
+
+
 class TestValidateAndFixThinkingParam:
     """Tests for validate_and_fix_thinking_param."""
 
@@ -6444,6 +6490,53 @@ async def test_acompletion_finishes_response_metadata_before_handing_the_respons
     assert snapshot["api_base"]
 
 
+class _GatedSyncLoggingHookRecorder(CustomLogger):
+    def __init__(self) -> None:
+        super().__init__()
+        self.seen: Final = queue.SimpleQueue[str | None]()
+        self.release: Final = threading.Event()
+
+    def logging_hook(
+        self, kwargs: dict[str, object], result: object, call_type: str
+    ) -> tuple[dict[str, object], object]:
+        self.seen.put(result.id if isinstance(result, litellm.ModelResponse) else None)
+        self.release.wait(timeout=5)
+        return kwargs, result
+
+
+@pytest.mark.asyncio
+async def test_acompletion_runs_a_custom_logger_sync_logging_hook_exactly_once(monkeypatch: pytest.MonkeyPatch) -> None:
+    def legacy_sync_callback(
+        kwargs: dict[str, object], response: litellm.ModelResponse, start_time: datetime, end_time: datetime
+    ) -> None:
+        pass
+
+    recorder: Final = _GatedSyncLoggingHookRecorder()
+    monkeypatch.setattr(litellm, "success_callback", [legacy_sync_callback, recorder])
+    logging_futures: Final = queue.SimpleQueue[Future[object]]()
+    real_submit: Final = logging_executor.submit
+
+    def submit_and_track(fn: Callable[..., object], *args: object, **kwargs: object) -> Future[object]:
+        future: Final = real_submit(fn, *args, **kwargs)
+        logging_futures.put(future)
+        return future
+
+    with patch(  # test-quality-ok: wraps the real submit only to collect the futures to join, the pool still runs
+        "litellm.litellm_core_utils.litellm_logging.executor.submit", side_effect=submit_and_track
+    ):
+        response: Final = await litellm.acompletion(
+            model="gpt-5.5",
+            messages=[{"role": "user", "content": "hi"}],
+            mock_response="Hello there!",
+            num_retries=0,
+        )
+        await asyncio.sleep(0)
+    recorder.release.set()
+    for _ in range(logging_futures.qsize()):
+        logging_futures.get_nowait().result(timeout=5)
+    assert [recorder.seen.get_nowait() for _ in range(recorder.seen.qsize())] == [response.id]
+
+
 def test_completion_finishes_response_metadata_before_handing_the_response_to_the_logging_thread():
     with _recording_hidden_params_at_submit("litellm.utils.executor.submit") as seen:
         litellm.completion(
@@ -6455,3 +6548,11 @@ def test_completion_finishes_response_metadata_before_handing_the_response_to_th
     assert snapshot["litellm_call_id"]
     assert snapshot["response_cost"] is not None
     assert snapshot["api_base"]
+
+
+def test_get_model_info_carries_cache_read_input_audio_token_cost(monkeypatch):
+    monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
+    monkeypatch.setattr(litellm, "model_cost", litellm.get_model_cost_map(url=""))
+    info = litellm.get_model_info("gpt-realtime-2.1-mini", custom_llm_provider="openai")
+    assert info["cache_read_input_audio_token_cost"] == 3e-07
+    assert info["cache_read_input_token_cost"] == 6e-08

tests/test_litellm_rust/ocr/test_lifecycle.py

@@ -806,7 +806,7 @@ async def test_shared_call_limits_still_reject_before_reading_ocr_file(
     monkeypatch.setattr(litellm, "_current_cost", 2)
     monkeypatch.setattr(litellm, "num_retries_per_request", 1 if limit == "retries" else None)
     expected: Final = litellm.BudgetExceededError if limit == "budget" else RuntimeError
-    arguments: Final = {"document": {"type": "file", "file": File()}, "metadata": {"previous_models": ["earlier"]}}
+    arguments: Final = {"document": {"type": "file", "file": File()}, "metadata": {"attempted_retries": 1}}
     with pytest.raises(expected, match=r"Budget has been exceeded|Max retries per request hit"):
         await call_aocr(ocr_server, **arguments) if asynchronous else call_ocr(ocr_server, **arguments)
     assert reads == []

ui/litellm-dashboard/src/components/ModelInfoEditForm.tsx

@@ -271,6 +271,7 @@ const displayCost = (localModelData: any, field: TouchedPricingField): string =>
 interface ModelInfoEditFormProps {
   localModelData: any;
   modelData: { model_info: { team_id?: string | null } & Record<string, unknown> };
+  teamAlias: string | null;
   accessToken: string | null;
   isEditing: boolean;
   isSaving: boolean;
@@ -341,6 +342,7 @@ const ChipList: React.FC<{ values: unknown; emptyLabel: string }> = ({ values, e
 const ModelInfoEditForm: React.FC<ModelInfoEditFormProps> = ({
   localModelData,
   modelData,
+  teamAlias,
   accessToken,
   isEditing,
   isSaving,
@@ -799,8 +801,12 @@ const ModelInfoEditForm: React.FC<ModelInfoEditFormProps> = ({
             </div>
 
             <div>
-              <FieldLabel>Team ID</FieldLabel>
-              <Display>{modelData.model_info.team_id || "Not Set"}</Display>
+              <FieldLabel>Team</FieldLabel>
+              <Display>
+                {teamAlias
+                  ? `${teamAlias} (${modelData.model_info.team_id})`
+                  : modelData.model_info.team_id || "Not Set"}
+              </Display>
             </div>
           </div>
 

ui/litellm-dashboard/src/components/model_info_view.test.tsx

@@ -42,6 +42,11 @@ vi.mock("@/app/(dashboard)/hooks/models/useModelCostMap", () => ({
   useModelCostMap: (...args: any[]) => mockUseModelCostMap(...args),
 }));
 
+const mockUseTeams = vi.fn();
+vi.mock("@/app/(dashboard)/hooks/teams/useTeams", () => ({
+  useTeams: () => mockUseTeams(),
+}));
+
 const mockUsePtuCostAttributionEnabled = vi.fn();
 vi.mock("@/app/(dashboard)/hooks/uiSettings/usePtuCostAttributionEnabled", () => ({
   usePtuCostAttributionEnabled: () => mockUsePtuCostAttributionEnabled(),
@@ -102,6 +107,7 @@ describe("ModelInfoView", () => {
     });
     vi.clearAllMocks();
     mockUsePtuCostAttributionEnabled.mockReturnValue(false);
+    mockUseTeams.mockReturnValue({ data: undefined, isLoading: false, error: null });
 
     mockUseModelsInfo.mockReturnValue({
       data: {
@@ -1305,6 +1311,78 @@ describe("ModelInfoView", () => {
     });
   });
 
+  describe("team alias", () => {
+    const teamModel = {
+      ...defaultModelData,
+      model_info: { ...defaultModelData.model_info, team_id: "team-1" },
+    };
+
+    beforeEach(() => {
+      mockUseModelsInfo.mockReturnValue({ data: { data: [teamModel] }, isLoading: false, error: null });
+      mockModelInfoV1Call.mockResolvedValue({ data: [teamModel] });
+    });
+
+    const readRawJson = async (user: ReturnType<typeof userEvent.setup>) => {
+      await user.click(await screen.findByRole("tab", { name: /raw json/i }));
+      const pre = await screen.findByText(/"model_name": "GPT-4"/, { selector: "pre" });
+      return JSON.parse(pre.textContent ?? "");
+    };
+
+    it("shows the team alias next to the team id and adds team_alias to the raw JSON", async () => {
+      mockUseTeams.mockReturnValue({
+        data: [
+          { team_id: "team-0", team_alias: "other" },
+          { team_id: "team-1", team_alias: "alpha" },
+        ],
+        isLoading: false,
+        error: null,
+      });
+      const user = userEvent.setup();
+      render(<ModelInfoView {...DEFAULT_ADMIN_PROPS} />, { wrapper });
+
+      expect(await screen.findByText("alpha (team-1)")).toBeInTheDocument();
+
+      const raw = await readRawJson(user);
+      expect(raw.model_info).toMatchObject({ team_id: "team-1", team_alias: "alpha" });
+      const keys = Object.keys(raw.model_info);
+      expect(keys.indexOf("team_alias")).toBe(keys.indexOf("team_id") + 1);
+    });
+
+    it("falls back to the bare team id when the team is not in the caller's team list", async () => {
+      mockUseTeams.mockReturnValue({
+        data: [{ team_id: "team-0", team_alias: "other" }],
+        isLoading: false,
+        error: null,
+      });
+      const user = userEvent.setup();
+      render(<ModelInfoView {...DEFAULT_ADMIN_PROPS} />, { wrapper });
+
+      expect(await screen.findByText("team-1")).toBeInTheDocument();
+
+      const raw = await readRawJson(user);
+      expect(raw.model_info.team_id).toBe("team-1");
+      expect(raw.model_info).not.toHaveProperty("team_alias");
+    });
+
+    it("shows Not Set and no team_alias for a model without a team", async () => {
+      mockUseModelsInfo.mockReturnValue({ data: { data: [defaultModelData] }, isLoading: false, error: null });
+      mockModelInfoV1Call.mockResolvedValue({ data: [defaultModelData] });
+      mockUseTeams.mockReturnValue({
+        data: [{ team_id: "team-1", team_alias: "alpha" }],
+        isLoading: false,
+        error: null,
+      });
+      const user = userEvent.setup();
+      render(<ModelInfoView {...DEFAULT_ADMIN_PROPS} />, { wrapper });
+
+      expect(await screen.findByText("Team")).toBeInTheDocument();
+      expect(screen.queryByText(/alpha/)).not.toBeInTheDocument();
+
+      const raw = await readRawJson(user);
+      expect(raw.model_info).not.toHaveProperty("team_alias");
+    });
+  });
+
   it("renders the provider card logo from the bundled provider map", async () => {
     render(<ModelInfoView {...DEFAULT_ADMIN_PROPS} />, { wrapper });
 

ui/litellm-dashboard/src/components/model_info_view.tsx

@@ -169,6 +169,12 @@ export default function ModelInfoView({
   // Keep modelData variable name for backwards compatibility
   const modelData = transformedModelData;
 
+  const teamAlias = teams?.find((team) => team.team_id === modelData?.model_info?.team_id)?.team_alias || null;
+  const rawModelInfoEntries = Object.entries(modelData?.model_info ?? {}).flatMap((entry) =>
+    entry[0] === "team_id" && teamAlias ? [entry, ["team_alias", teamAlias]] : [entry],
+  );
+  const rawModelData = modelData && { ...modelData, model_info: Object.fromEntries(rawModelInfoEntries) };
+
   const canEditModel = canModifyModel({ userRole, userID, isViewOnly }, teams ?? null, {
     teamId: modelData?.model_info?.team_id,
     isDbModel: modelData?.model_info?.db_model === true,
@@ -765,6 +771,7 @@ export default function ModelInfoView({
                 <ModelInfoEditForm
                   localModelData={localModelData}
                   modelData={modelData}
+                  teamAlias={teamAlias}
                   accessToken={accessToken}
                   isEditing={isEditing}
                   isSaving={isSaving}
@@ -788,7 +795,9 @@ export default function ModelInfoView({
 
           <TabsContent value="raw" keepMounted>
             <Card className="block p-6">
-              <pre className="bg-muted p-4 rounded-sm text-xs overflow-auto">{JSON.stringify(modelData, null, 2)}</pre>
+              <pre className="bg-muted p-4 rounded-sm text-xs overflow-auto">
+                {JSON.stringify(rawModelData, null, 2)}
+              </pre>
             </Card>
           </TabsContent>
         </div>

ui/litellm-dashboard/src/lib/http/schema.d.ts

@@ -8938,7 +8938,7 @@ export interface paths {
          *     `model_info.direct_access` when the proxy database is connected.
          *
          *     Returns:
-         *         Returns a dictionary containing information about each model.
+         *         A JSON response whose `data` list holds one entry per model.
          *
          *     Example Response:
          *     ```json
@@ -16781,7 +16781,6 @@ export interface paths {
          *     - permissions: Optional[dict] - [Not Implemented Yet] User-specific permissions, eg. turning off pii masking.
          *     - metadata: Optional[dict] - Metadata for user, store information for user. Example metadata = {"team": "core-infra", "app": "app2", "email": "ishaan@berri.ai" }
          *     - max_parallel_requests: Optional[int] - Rate limit a user based on the number of parallel requests. Raises 429 error, if user's parallel requests > x.
-         *     - soft_budget: Optional[float] - Get alerts when user crosses given budget, doesn't block requests.
          *     - model_max_budget: Optional[dict] - Model-specific max budget for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-budgets-to-keys)
          *     - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}.
          *     - model_rpm_limit: Optional[float] - Model-specific rpm limit for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-limits-to-keys)
@@ -16887,7 +16886,6 @@ export interface paths {
          *         - permissions: Optional[dict] - [Not Implemented Yet] User-specific permissions, eg. turning off pii masking.
          *         - metadata: Optional[dict] - Metadata for user, store information for user. Example metadata = {"team": "core-infra", "app": "app2", "email": "ishaan@berri.ai" }
          *         - max_parallel_requests: Optional[int] - Rate limit a user based on the number of parallel requests. Raises 429 error, if user's parallel requests > x.
-         *         - soft_budget: Optional[float] - Get alerts when user crosses given budget, doesn't block requests.
          *         - model_max_budget: Optional[dict] - Model-specific max budget for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-budgets-to-keys)
          *         - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}.
          *         - model_rpm_limit: Optional[float] - Model-specific rpm limit for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-limits-to-keys)
@@ -19086,7 +19084,7 @@ export interface paths {
          *     `model_info.direct_access` when the proxy database is connected.
          *
          *     Returns:
-         *         Returns a dictionary containing information about each model.
+         *         A JSON response whose `data` list holds one entry per model.
          *
          *     Example Response:
          *     ```json
@@ -25811,6 +25809,11 @@ export interface components {
              * @description If True, lets keys address Responses API ids that this proxy did not issue (raw provider ids, or ids issued before response-id encryption was configured). Such an id carries no owner, so no ownership check can run on it; ids this proxy did issue keep full ownership enforcement. Off by default, in which case an unrecognized response id is rejected with 403
              */
             allow_unmanaged_response_ids?: boolean | null;
+            /**
+             * Allowed File Extensions
+             * @description the only file extensions (e.g. ['.jsonl', '.pdf', '.txt']) accepted on /v1/files uploads, for any purpose, matched case-insensitively against the uploaded filename. Files with any other extension, or none, are rejected. An empty list rejects every upload. Unset means no allowlist is applied
+             */
+            allowed_file_extensions?: string[] | null;
             /**
              * Allowed Routes
              * @description Proxy API Endpoints you want users to be able to access
@@ -25833,7 +25836,7 @@ export interface components {
             background_health_checks?: boolean | null;
             /**
              * Blocked File Extensions
-             * @description file extensions (e.g. ['.exe', '.sh']) rejected on /v1/files uploads, for any purpose, matched case-insensitively against the uploaded filename
+             * @description file extensions (e.g. ['.exe', '.sh']) rejected on /v1/files uploads, for any purpose, matched case-insensitively against the uploaded filename. Deprecated in favour of allowed_file_extensions; still enforced, after the allowlist, when set
              */
             blocked_file_extensions?: string[] | null;
             /**