Merge 812f0054a4 into e1580b7b83

litellm/llms/azure/chat/gpt_reasoning_transformation.py

@@ -24,6 +24,7 @@ class AzureOpenAIGPT5Config(AzureOpenAIConfig, OpenAIGPT5Config):
 
     GPT5_SERIES_ROUTE = "gpt5_series/"
     OPT_IN_REASONING_EFFORTS: ClassVar[frozenset[str]] = frozenset(("xhigh", "none"))
+    LOOKS_UP_BASE_MODEL: ClassVar[bool] = True
 
     @classmethod
     def _model_map_lookup_name(cls, model: str) -> str:
@@ -88,7 +89,7 @@ class AzureOpenAIGPT5Config(AzureOpenAIConfig, OpenAIGPT5Config):
         drop_params: bool,
         api_version: str = "",
         *,
-        declared_reasoning_efforts: Collection[str] = (),
+        declared_reasoning_efforts: Collection[str] | None = None,
     ) -> dict:
         return OpenAIGPT5Config.map_openai_params(
             self,

litellm/llms/azure_ai/chat/transformation.py

@@ -106,7 +106,7 @@ class AzureAIStudioConfig(OpenAIConfig):
         model: str,
         drop_params: bool,
         *,
-        declared_reasoning_efforts: Collection[str] = (),
+        declared_reasoning_efforts: Collection[str] | None = None,
     ) -> dict[str, object]:  # mutable-ok: OpenAIConfig.map_openai_params signature
         if not azureAIGPT5Config.is_model_gpt_5_model(model):
             return super().map_openai_params(

litellm/llms/bedrock/chat/converse_transformation.py

@@ -117,6 +117,7 @@ from litellm.utils import (
 )
 
 from ..common_utils import (
+    DROP_DISABLED_REASONING_EFFORT_WARNING,
     BedrockError,
     BedrockModelInfo,
     bedrock_cache_point_ttl_supported,
@@ -1192,11 +1193,7 @@ class AmazonConverseConfig(BaseConfig):
                         ),
                         status_code=400,
                     )
-                verbose_logger.debug(
-                    "Dropping unsupported `reasoning_effort=%s` for Bedrock model=%s.",
-                    value,
-                    model,
-                )
+                verbose_logger.warning(DROP_DISABLED_REASONING_EFFORT_WARNING, "reasoning_effort", value, model)
             elif param == "reasoning_effort" and isinstance(value, str):
                 self._handle_reasoning_effort_parameter(
                     model=model, reasoning_effort=value, optional_params=optional_params

litellm/llms/bedrock/common_utils.py

@@ -1314,6 +1314,12 @@ def _mantle_api_base_from_env() -> str | None:
     return next((base[: -len(suffix)] for suffix in _MANTLE_OPENAI_BASE_SUFFIXES if base.endswith(suffix)), base)
 
 
+DROP_DISABLED_REASONING_EFFORT_WARNING: Final = (
+    "Dropping %s=%r for Bedrock model=%r (drop_params=True) because the model map marks that level unsupported. "
+    "Send a level the model supports."
+)
+
+
 def bedrock_reasoning_effort_disabled(model: str, effort: str) -> bool:
     from litellm.utils import is_explicitly_disabled_factory
 

litellm/llms/bedrock/responses/transformation.py

@@ -52,6 +52,7 @@ from litellm.llms.base_llm.responses.codex_compat import drop_unsupported_tools,
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
 from litellm.llms.bedrock.common_utils import (
     BEDROCK_CHAT_COMPLETIONS_ROUTE_PREFIX,
+    DROP_DISABLED_REASONING_EFFORT_WARNING,
     BedrockError,
     bedrock_reasoning_effort_disabled,
     bedrock_supports_openai_responses,
@@ -173,7 +174,7 @@ def _without_disabled_reasoning_effort(
             ),
             status_code=400,
         )
-    verbose_logger.debug("Dropping unsupported `reasoning.effort=%s` for Bedrock model=%s.", effort, model)
+    verbose_logger.warning(DROP_DISABLED_REASONING_EFFORT_WARNING, "reasoning.effort", effort, model)
     rest: Final = {key: value for key, value in reasoning.items() if key != "effort"}
     without_reasoning: Final = {key: value for key, value in params.items() if key != "reasoning"}
     return {**without_reasoning, "reasoning": rest} if rest else without_reasoning

litellm/llms/openai/chat/gpt_reasoning_transformation.py

@@ -5,8 +5,10 @@ from collections.abc import Collection
 from typing import ClassVar, Final
 
 import litellm
+from litellm._logging import verbose_logger
 from litellm.utils import (
     declared_value_factory,
+    has_catalog_entry,
     is_explicitly_disabled_factory,
     supports_factory,
 )
@@ -16,6 +18,25 @@ from .gpt_transformation import OpenAIGPTConfig
 _GPT_SERIES_VERSION: Final = re.compile(r"^gpt-(\d+)(?:\.(\d+))?(?=[.-]|$)")
 
 
+def refused_reasoning_effort_warning(*, suggest_declared_levels: bool, suggest_base_model: bool) -> str:
+    return "".join(
+        (
+            "Dropping %s=%r for model=%r (drop_params=True) because the model map does not mark that level "
+            "supported for %r.",
+            (
+                " If the deployment supports it, list every level it supports in model_info.reasoning_effort_levels."
+                if suggest_declared_levels
+                else ""
+            ),
+            (
+                " If the deployment name is not its underlying model, set model_info.base_model to that model."
+                if suggest_base_model
+                else ""
+            ),
+        )
+    )
+
+
 def _catalogue_declares_default_effort() -> bool:
     """Whether the loaded cost map carries default_reasoning_effort for ANY entry.
 
@@ -93,6 +114,7 @@ class OpenAIGPT5Config(OpenAIGPTConfig):
 
     OPT_IN_REASONING_EFFORTS: ClassVar[frozenset[str]] = frozenset(("xhigh",))
     ROUTE_SPECIFIC_REASONING_EFFORTS: ClassVar[frozenset[str]] = frozenset(("max",))
+    LOOKS_UP_BASE_MODEL: ClassVar[bool] = False
 
     @classmethod
     def is_model_gpt_5_model(cls, model: str) -> bool:
@@ -307,7 +329,7 @@ class OpenAIGPT5Config(OpenAIGPTConfig):
         model: str,
         drop_params: bool,
         *,
-        declared_reasoning_efforts: Collection[str] = (),
+        declared_reasoning_efforts: Collection[str] | None = None,
     ) -> dict:
         if self.is_model_gpt_5_search_model(model):
             if "max_tokens" in non_default_params:
@@ -338,10 +360,21 @@ class OpenAIGPT5Config(OpenAIGPTConfig):
 
         if (
             requested_effort is not None
-            and requested_effort not in declared_reasoning_efforts
+            and requested_effort not in (declared_reasoning_efforts or ())
             and self.refuses_reasoning_effort_level(model, requested_effort)
         ):
             if litellm.drop_params or drop_params:
+                verbose_logger.warning(
+                    refused_reasoning_effort_warning(
+                        suggest_declared_levels=declared_reasoning_efforts is not None,
+                        suggest_base_model=self.LOOKS_UP_BASE_MODEL
+                        and not has_catalog_entry(self._model_map_lookup_name(model), None),
+                    ),
+                    "reasoning_effort",
+                    requested_effort,
+                    model,
+                    self._model_map_lookup_name(model),
+                )
                 non_default_params.pop("reasoning_effort", None)
                 optional_params.pop("reasoning_effort", None)
             else:

litellm/llms/openai/responses/transformation.py

@@ -27,6 +27,7 @@ from litellm.litellm_core_utils.url_utils import encode_url_path_segment, join_a
 from litellm.llms.base_llm.responses.transformation import BaseResponsesAPIConfig
 from litellm.llms.openai.chat.gpt_reasoning_transformation import (
     is_gpt_reasoning_series_name,
+    refused_reasoning_effort_warning,
     unsupported_reasoning_effort_message,
 )
 from litellm.responses.litellm_completion_transformation.custom_tools import TOOL_CALL_ITEM_ID_PREFIX_BY_TYPE
@@ -50,6 +51,9 @@ from ..common_utils import OpenAIError
 from ..workload_identity import get_workload_identity_bearer_token, resolve_openai_workload_identity_config
 
 OPENAI_RESPONSES_API_MIN_MAX_OUTPUT_TOKENS: Final = 16
+DROP_UNSUPPORTED_REASONING_WARNING: Final = (
+    "Dropping reasoning=%r for model=%r (drop_params=True) because its model cost map entry lacks supports_reasoning."
+)
 _RAW_RESPONSE_JSON: Final = TypeAdapter(dict[str, object], config=ConfigDict(strict=True))
 
 if TYPE_CHECKING:
@@ -324,12 +328,19 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
             return declared
         return info["key"] in _bundled_openai_reasoning_models()
 
+    @staticmethod
+    def _is_known_non_reasoning_model(model: str) -> bool:
+        from litellm.utils import is_known_non_reasoning_model
+
+        return is_known_non_reasoning_model(model.split("/")[-1], LlmProviders.OPENAI.value)
+
     @staticmethod
     def _requests_reasoning_effort(reasoning: object) -> bool:
-        effort: Final = (
-            reasoning.get("effort") if isinstance(reasoning, Mapping) else getattr(reasoning, "effort", None)
-        )
-        return effort is not None
+        return OpenAIResponsesAPIConfig._requested_effort(reasoning) is not None
+
+    @staticmethod
+    def _requested_effort(reasoning: object) -> object:
+        return reasoning.get("effort") if isinstance(reasoning, Mapping) else getattr(reasoning, "effort", None)
 
     @staticmethod
     def _enforce_min_max_output_tokens(max_output_tokens: "int | None") -> "int | None":
@@ -371,17 +382,20 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
     ) -> dict:
         return dict(
             self.map_openai_params_for_deployment(
-                response_api_optional_params, model, drop_params, declared_reasoning_efforts=()
+                response_api_optional_params, model, drop_params, declared_reasoning_efforts=None
             )
         )
 
+    def honors_declared_reasoning_efforts(self) -> bool:
+        return type(self).map_openai_params is OpenAIResponsesAPIConfig.map_openai_params
+
     def map_openai_params_for_deployment(
         self,
         response_api_optional_params: ResponsesAPIOptionalRequestParams,
         model: str,
         drop_params: bool,
         *,
-        declared_reasoning_efforts: Collection[str],
+        declared_reasoning_efforts: Collection[str] | None,
     ) -> Mapping[str, object]:
         """No mapping applied since inputs are in OpenAI spec already.
 
@@ -401,7 +415,9 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
             and not self._supports_reasoning_param(model=model)
         ):
             if drop_params or litellm.drop_params:
-                params.pop("reasoning", None)
+                dropped_reasoning: Final = params.pop("reasoning", None)
+                if self._requested_effort(dropped_reasoning) != "none" or not self._is_known_non_reasoning_model(model):
+                    verbose_logger.warning(DROP_UNSUPPORTED_REASONING_WARNING, dropped_reasoning, model)
             else:
                 raise litellm.UnsupportedParamsError(
                     message=(
@@ -419,7 +435,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
             if (
                 isinstance(reasoning, dict)
                 and isinstance(requested_effort, str)
-                and requested_effort not in declared_reasoning_efforts
+                and requested_effort not in (declared_reasoning_efforts or ())
                 and self._effort_level_is_disabled(lookup_name, requested_effort)
             ):
                 if not (drop_params or litellm.drop_params):
@@ -427,6 +443,15 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
                         message=unsupported_reasoning_effort_message("reasoning.effort", requested_effort, model),
                         status_code=400,
                     )
+                verbose_logger.warning(
+                    refused_reasoning_effort_warning(
+                        suggest_declared_levels=declared_reasoning_efforts is not None, suggest_base_model=False
+                    ),
+                    "reasoning.effort",
+                    requested_effort,
+                    model,
+                    lookup_name,
+                )
                 remaining: Final = reasoning.copy()
                 remaining.pop("effort", None)
                 if remaining:

litellm/responses/utils.py

@@ -131,14 +131,14 @@ def _provider_mapped_params(
     params: ResponsesAPIOptionalRequestParams,
     model: str,
     drop_params: bool,
-    declared_reasoning_efforts: Collection[str],
+    declared_reasoning_efforts: Collection[str] | None,
 ) -> Mapping[str, object]:
     from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
 
     if (
-        declared_reasoning_efforts
+        declared_reasoning_efforts is not None
         and isinstance(config, OpenAIResponsesAPIConfig)
-        and type(config).map_openai_params is OpenAIResponsesAPIConfig.map_openai_params
+        and config.honors_declared_reasoning_efforts()
     ):
         return config.map_openai_params_for_deployment(
             params, model, drop_params, declared_reasoning_efforts=declared_reasoning_efforts
@@ -287,7 +287,7 @@ class ResponsesAPIRequestUtils:
         response_api_optional_params: ResponsesAPIOptionalRequestParams,
         allowed_openai_params: list[str] | None = None,
         drop_params: bool | None = None,
-        declared_reasoning_efforts: Collection[str] = (),
+        declared_reasoning_efforts: Collection[str] | None = None,
     ) -> dict:
         """
         Get optional parameters for the responses API.

litellm/router_utils/reasoning_effort_capability.py

@@ -426,6 +426,27 @@ def _log_remap(
     logged_deployment_id: Final = _single_line(deployment_id)
     logged_model_group: Final = _single_line(model_group)
     logged_deployment_model: Final = _single_line(deployment_model)
+    extra: Final = MappingProxyType(
+        {
+            "reasoning_effort_deployment_id": logged_deployment_id,
+            "reasoning_effort_model_group": logged_model_group,
+            "reasoning_effort_deployment_model": logged_deployment_model,
+            "reasoning_effort_requested": logged_requested_level,
+            "reasoning_effort_sent": logged_sent_level,
+            "reasoning_effort_rounding": rounding,
+        }
+    )
+    if sent_level is None and requested_level != _DEFAULT_EFFORT:
+        verbose_logger.warning(
+            "Dropping reasoning level %s (drop_params=True) for deployment %s, model group %s, model %s because its "
+            "model_info.reasoning_effort_levels does not list that level. List every level the deployment supports.",
+            logged_requested_level,
+            logged_deployment_id,
+            logged_model_group,
+            logged_deployment_model,
+            extra=extra,
+        )
+        return
     verbose_logger.info(
         "reasoning_effort %s sent as %s (rounding %s) to deployment %s, model group %s, model %s",
         logged_requested_level,
@@ -434,14 +455,7 @@ def _log_remap(
         logged_deployment_id,
         logged_model_group,
         logged_deployment_model,
-        extra={
-            "reasoning_effort_deployment_id": logged_deployment_id,
-            "reasoning_effort_model_group": logged_model_group,
-            "reasoning_effort_deployment_model": logged_deployment_model,
-            "reasoning_effort_requested": logged_requested_level,
-            "reasoning_effort_sent": logged_sent_level,
-            "reasoning_effort_rounding": rounding,
-        },
+        extra=extra,
     )
 
 

litellm/utils.py

@@ -91,7 +91,7 @@ from litellm.litellm_core_utils.fallback_generalizations import (
     match_capability_generalizations,
     match_fill_missing_generalizations,
 )
-from litellm.litellm_core_utils.get_model_cost_map import AliasedCostEntry
+from litellm.litellm_core_utils.get_model_cost_map import AliasedCostEntry, GetModelCostMap
 from litellm.litellm_core_utils.model_param_helper import peek_reasoning_summary_aliases
 from litellm.litellm_core_utils.offline_mode import is_offline_mode
 from litellm.litellm_core_utils.sensitive_data_masker import redact_credentials_in_payload
@@ -4974,6 +4974,46 @@ def _declared_reasoning_efforts(model_info: Mapping[str, object] | None) -> tupl
     return deployment_reasoning_effort_levels(model_info) or ()
 
 
+_DROPPED_REASONING_PARAMS: Final = ("reasoning_effort", "thinking")
+
+
+def _turns_reasoning_off(value: object) -> bool:
+    match value:
+        case "none" | {"effort": "none"} | {"type": "disabled"}:
+            return True
+        case _:
+            return False
+
+
+def is_known_non_reasoning_model(model: str, custom_llm_provider: str | None) -> bool:
+    info: Final = lookup_model_info_for_registration(model, custom_llm_provider)
+    if isinstance(info, ModelInfoLookupMiss) or info.get("supports_reasoning") is True:
+        return False
+    return info.get("supports_reasoning") is False or info["key"] in GetModelCostMap.loaded_model_cost_map()
+
+
+def has_catalog_entry(model: str, custom_llm_provider: str | None) -> bool:
+    info: Final = lookup_model_info_for_registration(model, custom_llm_provider)
+    return not isinstance(info, ModelInfoLookupMiss) and info["key"] in GetModelCostMap.loaded_model_cost_map()
+
+
+def _dropped_reasoning_warning(model: str, custom_llm_provider: str | None) -> str:
+    return "".join(
+        (
+            "Dropping %s=%r for model=%r (drop_params=True) because %s does not support %s for it.",
+            (
+                ""
+                if has_catalog_entry(model, custom_llm_provider)
+                else " If the deployment is a reasoning model, set model_info.base_model to its underlying model."
+            ),
+        )
+    )
+
+
+def _warns_on_dropped_reasoning(value: object, model: str, custom_llm_provider: str | None) -> bool:
+    return not (_turns_reasoning_off(value) and is_known_non_reasoning_model(model, custom_llm_provider))
+
+
 def get_optional_params(
     # use the openai defaults
     # https://platform.openai.com/docs/api-reference/chat/create
@@ -5061,19 +5101,30 @@ def get_optional_params(
         verbose_logger.info("\nLiteLLM completion() model= %s; provider = %s", model, custom_llm_provider)
         verbose_logger.debug("\nLiteLLM: Params passed to completion() %s", passed_params)
         verbose_logger.debug("\nLiteLLM: Non-Default params passed to completion() %s", non_default_params)
-        unsupported_params: Final = {}
-        for k in non_default_params:
-            if k not in supported_params:
-                if k in PROVIDER_UNVALIDATED_PARAMS:
-                    continue
-                if k == "n" and n == 1:  # langchain sends n=1 as a default value
-                    continue  # skip this param
-                unsupported_params[k] = non_default_params[k]
+        unsupported_params: Final[MappingProxyType[str, object]] = MappingProxyType(
+            {
+                k: non_default_params[k]
+                for k in non_default_params
+                if k not in supported_params and k not in PROVIDER_UNVALIDATED_PARAMS and not (k == "n" and n == 1)
+            }
+        )
 
         if unsupported_params:
             if litellm.drop_params is True or (drop_params is not None and drop_params is True):
                 for k in unsupported_params:
                     non_default_params.pop(k, None)
+                for reasoning_param in _DROPPED_REASONING_PARAMS:
+                    if reasoning_param in unsupported_params and _warns_on_dropped_reasoning(
+                        unsupported_params[reasoning_param], base_model or model, custom_llm_provider
+                    ):
+                        verbose_logger.warning(
+                            _dropped_reasoning_warning(base_model or model, custom_llm_provider),
+                            reasoning_param,
+                            unsupported_params[reasoning_param],
+                            model,
+                            custom_llm_provider,
+                            reasoning_param,
+                        )
             else:
                 raise UnsupportedParamsError(
                     status_code=500,

tests/unit/llms/azure/chat/test_azure_gpt_reasoning_transformation.py

@@ -1,3 +1,4 @@
+import logging
 from collections.abc import Mapping
 from typing import Final
 
@@ -279,6 +280,29 @@ def test_azure_gpt5_reasoning_effort_none_dropped(config: AzureOpenAIGPT5Config)
     assert "reasoning_effort" not in params or params.get("reasoning_effort") != "none"
 
 
+@pytest.mark.parametrize(("model", "names_base_model"), [("azure/litellm-qa-sol-alias", True), ("azure/gpt-5", False)])
+def test_azure_dropped_effort_warning_suggests_base_model_only_for_a_name_the_map_lacks(
+    config: AzureOpenAIGPT5Config, caplog: pytest.LogCaptureFixture, model: str, names_base_model: bool
+) -> None:
+    with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+        params: Final = config.map_openai_params(
+            non_default_params={"reasoning_effort": "none"},
+            optional_params={},
+            model=model,
+            drop_params=True,
+            api_version="2025-04-01-preview",
+            declared_reasoning_efforts=(),
+        )
+
+    assert "reasoning_effort" not in params
+    drop_warnings: Final = [
+        record.getMessage() for record in caplog.records if "Dropping reasoning_effort='none'" in record.getMessage()
+    ]
+    assert len(drop_warnings) == 1
+    assert "model_info.reasoning_effort_levels" in drop_warnings[0]
+    assert ("model_info.base_model" in drop_warnings[0]) is names_base_model
+
+
 # Logprobs support tests for Azure GPT-5.2
 def test_azure_gpt5_2_supports_logprobs(config: AzureOpenAIGPT5Config):
     """Test that Azure GPT-5.2 models support logprobs parameters.

tests/unit/llms/bedrock/chat/test_converse_transformation.py

@@ -1,6 +1,7 @@
 import base64
 import copy
 import json
+import logging
 import os
 import re
 from collections.abc import Iterator, Mapping
@@ -966,7 +967,9 @@ def test_reasoning_effort_maps_to_reasoning_effort_for_openai_gpt5_converse(mode
         "us.openai.gpt-6-astra",
     ],
 )
-def test_openai_gpt5_converse_rejects_effort_level_disabled_in_model_map(model, local_model_cost_map):
+def test_openai_gpt5_converse_rejects_effort_level_disabled_in_model_map(
+    model: str, local_model_cost_map: None, caplog: pytest.LogCaptureFixture
+) -> None:
     config = AmazonConverseConfig()
     assert litellm.utils.is_explicitly_disabled_factory(
         model=model, custom_llm_provider="bedrock_converse", key="supports_minimal_reasoning_effort"
@@ -980,15 +983,19 @@ def test_openai_gpt5_converse_rejects_effort_level_disabled_in_model_map(model,
             drop_params=False,
         )
 
-    optional_params = config.map_openai_params(
-        non_default_params={"reasoning_effort": "minimal"},
-        optional_params={},
-        model=model,
-        drop_params=True,
-    )
+    with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+        optional_params: Final = config.map_openai_params(
+            non_default_params={"reasoning_effort": "minimal"},
+            optional_params={},
+            model=model,
+            drop_params=True,
+        )
     _, additional_request_params, _, _ = config._prepare_request_params(optional_params, model)
     assert "reasoning" not in additional_request_params
     assert "thinking" not in additional_request_params
+    dropped: Final = [r.getMessage() for r in caplog.records if "Dropping reasoning_effort='minimal'" in r.getMessage()]
+    assert len(dropped) == 1 and model in dropped[0]
+    assert "model_info" not in dropped[0]
 
 
 @pytest.mark.parametrize(

tests/unit/llms/bedrock/responses/test_bedrock_openai_responses.py

@@ -7,13 +7,17 @@ falls back to the Chat Completions bridge and rides Converse.
 import json
 import logging
 from importlib.resources import files
+from typing import Final
 from unittest.mock import patch
 
 import pytest
 
 import litellm
 from litellm.llms.bedrock.common_utils import bedrock_supports_openai_responses
-from litellm.llms.bedrock.responses.transformation import BedrockOpenAIResponsesConfig
+from litellm.llms.bedrock.responses.transformation import (
+    BedrockOpenAIResponsesConfig,
+    _without_disabled_reasoning_effort,
+)
 from litellm.responses.file_search.emulated_handler import should_use_emulated_file_search
 from litellm.types.router import GenericLiteLLMParams
 from litellm.types.utils import LlmProviders
@@ -337,13 +341,37 @@ class TestDisabledReasoningEffort:
             )
 
     @pytest.mark.parametrize("model", ["us.openai.gpt-5.6-luna", MODEL])
-    def test_effort_level_disabled_in_model_map_is_dropped_with_drop_params(self, model, local_model_cost_map):
-        params = _cfg().map_openai_params(
-            response_api_optional_params={"reasoning": {"effort": "minimal", "summary": "auto"}, "max_output_tokens": 64},
-            model=model,
-            drop_params=True,
-        )
+    def test_effort_level_disabled_in_model_map_is_dropped_with_a_warning(
+        self, model: str, local_model_cost_map: None, caplog: pytest.LogCaptureFixture
+    ) -> None:
+        with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+            params: Final = _cfg().map_openai_params(
+                response_api_optional_params={
+                    "reasoning": {"effort": "minimal", "summary": "auto"},
+                    "max_output_tokens": 64,
+                },
+                model=model,
+                drop_params=True,
+            )
         assert params == {"reasoning": {"summary": "auto"}, "max_output_tokens": 64}
+        dropped: Final = [
+            r.getMessage() for r in caplog.records if "Dropping reasoning.effort='minimal'" in r.getMessage()
+        ]
+        assert len(dropped) == 1 and model in dropped[0]
+        assert "model_info" not in dropped[0]
+
+    def test_bedrock_disabled_effort_drop_warns_on_its_own(
+        self, local_model_cost_map: None, caplog: pytest.LogCaptureFixture
+    ) -> None:
+        with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+            params: Final = _without_disabled_reasoning_effort(
+                {"reasoning": {"effort": "minimal"}, "max_output_tokens": 64}, MODEL, True
+            )
+        assert params == {"max_output_tokens": 64}
+        dropped: Final = [
+            r.getMessage() for r in caplog.records if "Dropping reasoning.effort='minimal'" in r.getMessage()
+        ]
+        assert len(dropped) == 1 and MODEL in dropped[0]
 
     def test_effort_only_reasoning_is_removed_when_dropped(self, local_model_cost_map):
         params = _cfg().map_openai_params(

tests/unit/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py

@@ -9,6 +9,7 @@ gate, the URL construction for both paths, and the shared Bearer auth.
 
 import copy
 import logging
+from typing import Final
 
 import pytest
 from botocore.exceptions import (
@@ -572,6 +573,32 @@ class TestBedrockMantleServiceTier:
         assert "priority" in str(mock_warning.call_args)
 
 
+class TestBedrockMantleRefusedEffort:
+    def test_dropping_a_refused_effort_does_not_suggest_declaring_levels_mantle_ignores(
+        self, local_cost_map: None, caplog: pytest.LogCaptureFixture
+    ) -> None:
+        refused: Final = next(
+            level
+            for level in ("none", "minimal", "low")
+            if litellm.model_cost["gpt-6.1-sol"].get(f"supports_{level}_reasoning_effort") is False
+        )
+        with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+            params: Final = BedrockMantleResponsesAPIConfig().map_openai_params(
+                response_api_optional_params={"reasoning": {"effort": refused}},
+                model="openai.gpt-6.1-sol",
+                drop_params=True,
+            )
+
+        assert "reasoning" not in params
+        drop_warnings: Final = [
+            record.getMessage()
+            for record in caplog.records
+            if f"Dropping reasoning.effort={refused!r}" in record.getMessage()
+        ]
+        assert len(drop_warnings) == 1
+        assert "model_info" not in drop_warnings[0]
+
+
 class TestBedrockMantleReasoningSummary:
     @pytest.mark.parametrize("summary", ["concise", "detailed"])
     def test_unsupported_reasoning_summary_dropped_when_drop_params_true(self, summary):

tests/unit/llms/openai/responses/test_openai_responses_transformation.py

@@ -1,4 +1,5 @@
 import json
+import logging
 from collections.abc import Callable, Mapping
 from types import SimpleNamespace
 from typing import Final, cast
@@ -15,6 +16,7 @@ from litellm.llms.azure.responses.transformation import AzureOpenAIResponsesAPIC
 from litellm.llms.base_llm.responses.transformation import BaseResponsesAPIConfig
 from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
 from litellm.responses.litellm_completion_transformation.transformation import LiteLLMCompletionResponsesConfig
+from litellm.responses.utils import ResponsesAPIRequestUtils
 from litellm.types.llms.openai import (
     ImageGenerationPartialImageEvent,
     IncompleteDetails,
@@ -2387,6 +2389,105 @@ class TestResponsesSurfaceSharesTheEffortRule:
         )
         assert mapped["reasoning"] == {"summary": "auto"}
 
+    @pytest.mark.parametrize(
+        ("declared_reasoning_efforts", "suggests_declared_levels"),
+        [((), True), (None, False)],
+        ids=["responses route reads model_info levels", "compact route reads none"],
+    )
+    def test_dropping_a_refused_effort_logs_one_warning_naming_only_the_fix_its_route_honors(
+        self,
+        local_model_cost_map: None,
+        caplog: pytest.LogCaptureFixture,
+        declared_reasoning_efforts: tuple[str, ...] | None,
+        suggests_declared_levels: bool,
+    ) -> None:
+        refused: Final = next(
+            level
+            for level in ("none", "minimal", "low")
+            if litellm.model_cost["gpt-6.1-sol"].get(f"supports_{level}_reasoning_effort") is False
+        )
+        with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+            mapped: Final = ResponsesAPIRequestUtils.get_optional_params_responses_api(
+                model="gpt-6.1-sol",
+                responses_api_provider_config=OpenAIResponsesAPIConfig(),
+                response_api_optional_params={"reasoning": {"effort": refused}},
+                drop_params=True,
+                declared_reasoning_efforts=declared_reasoning_efforts,
+            )
+
+        assert "reasoning" not in mapped
+        drop_warnings: Final = [record.getMessage() for record in caplog.records if "Dropping" in record.getMessage()]
+        assert len(drop_warnings) == 1
+        assert repr(refused) in drop_warnings[0] and "gpt-6.1-sol" in drop_warnings[0]
+        assert ("model_info.reasoning_effort_levels" in drop_warnings[0]) is suggests_declared_levels
+        assert "model_info.base_model" not in drop_warnings[0]
+
+    def test_dropping_reasoning_for_a_model_without_supports_reasoning_logs_one_warning(
+        self, local_model_cost_map: None, caplog: pytest.LogCaptureFixture
+    ) -> None:
+        with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+            mapped: Final = OpenAIResponsesAPIConfig().map_openai_params(
+                response_api_optional_params={"reasoning": {"effort": "medium"}},
+                model="gpt-4o",
+                drop_params=True,
+            )
+
+        assert "reasoning" not in mapped
+        drop_warnings: Final = [record.getMessage() for record in caplog.records if "Dropping" in record.getMessage()]
+        assert len(drop_warnings) == 1
+        assert all(fragment in drop_warnings[0] for fragment in ("'medium'", "gpt-4o", "supports_reasoning"))
+
+    def test_turning_reasoning_off_for_a_model_without_supports_reasoning_drops_it_silently(
+        self, local_model_cost_map: None, caplog: pytest.LogCaptureFixture
+    ) -> None:
+        with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+            mapped: Final = OpenAIResponsesAPIConfig().map_openai_params(
+                response_api_optional_params={"reasoning": {"effort": "none", "summary": "auto"}},
+                model="gpt-4o",
+                drop_params=True,
+            )
+
+        assert "reasoning" not in mapped
+        assert not [record for record in caplog.records if "Dropping" in record.getMessage()]
+
+    def test_turning_reasoning_off_warns_when_only_custom_pricing_describes_the_model(
+        self, local_model_cost_map: None, caplog: pytest.LogCaptureFixture
+    ) -> None:
+        litellm.register_model(
+            {
+                "pricing-only-test-model": {
+                    "litellm_provider": "openai",
+                    "mode": "responses",
+                    "input_cost_per_token": 0.000001,
+                    "output_cost_per_token": 0.000002,
+                }
+            }
+        )
+        with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+            mapped: Final = OpenAIResponsesAPIConfig().map_openai_params(
+                response_api_optional_params={"reasoning": {"effort": "none"}},
+                model="pricing-only-test-model",
+                drop_params=True,
+            )
+
+        assert "reasoning" not in mapped
+        drop_warnings: Final = [record.getMessage() for record in caplog.records if "Dropping" in record.getMessage()]
+        assert len(drop_warnings) == 1
+        assert all(fragment in drop_warnings[0] for fragment in ("'none'", "pricing-only-test-model"))
+
+    def test_a_forwarded_effort_logs_no_drop_warning(
+        self, local_model_cost_map: None, caplog: pytest.LogCaptureFixture
+    ) -> None:
+        with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+            mapped: Final = OpenAIResponsesAPIConfig().map_openai_params(
+                response_api_optional_params={"reasoning": {"effort": "medium"}},
+                model="gpt-6.1-sol",
+                drop_params=True,
+            )
+
+        assert mapped["reasoning"] == {"effort": "medium"}
+        assert not [record for record in caplog.records if "Dropping" in record.getMessage()]
+
     @pytest.mark.parametrize("level", ["none", "minimal", "low", "medium", "high", "xhigh", "max"])
     def test_gpt5_3_codex_forwards_exactly_the_effort_levels_its_map_row_allows(self, local_model_cost_map, level):
         """The gate turns a stale false into a local refusal, so the Responses-only codex row has to

tests/unit/llms/openai/test_gpt_reasoning_transformation.py

@@ -1,3 +1,4 @@
+import logging
 import re
 from collections.abc import Mapping
 from typing import Final
@@ -1589,3 +1590,55 @@ def test_gpt_6_astra_takes_the_reasoning_series_request_shape():
     assert "max_tokens" not in params
     assert params["reasoning_effort"] == "max"
     assert params["verbosity"] == "low"
+
+
+@pytest.mark.parametrize(
+    ("custom_llm_provider", "level", "advice"),
+    [
+        ("azure", "none", frozenset(("model_info.reasoning_effort_levels", "model_info.base_model"))),
+        ("azure", "xhigh", frozenset(("model_info.reasoning_effort_levels", "model_info.base_model"))),
+        ("openai", "xhigh", frozenset(("model_info.reasoning_effort_levels",))),
+        ("azure_ai", "xhigh", frozenset(("model_info.reasoning_effort_levels",))),
+        ("github_copilot", "xhigh", frozenset()),
+    ],
+)
+def test_a_reasoning_effort_the_level_gate_drops_logs_a_warning_naming_only_the_fixes_its_route_honors(
+    caplog: pytest.LogCaptureFixture, custom_llm_provider: str, level: str, advice: frozenset[str]
+) -> None:
+    with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+        params: Final = litellm.get_optional_params(
+            model="gpt-5.6-terra-ptu",
+            custom_llm_provider=custom_llm_provider,
+            reasoning_effort=level,
+            drop_params=True,
+        )
+
+    assert "reasoning_effort" not in params
+    drop_warnings: Final = [
+        record.getMessage()
+        for record in caplog.records
+        if record.levelno == logging.WARNING and "reasoning_effort" in record.getMessage()
+    ]
+    assert len(drop_warnings) == 1
+    assert repr(level) in drop_warnings[0] and "gpt-5.6-terra-ptu" in drop_warnings[0]
+    assert (
+        frozenset(
+            fragment
+            for fragment in ("model_info.reasoning_effort_levels", "model_info.base_model")
+            if fragment in drop_warnings[0]
+        )
+        == advice
+    )
+
+
+def test_a_reasoning_effort_the_level_gate_forwards_logs_no_drop_warning(caplog: pytest.LogCaptureFixture) -> None:
+    with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+        params: Final = litellm.get_optional_params(
+            model="gpt-5.6-terra-ptu",
+            custom_llm_provider="azure",
+            reasoning_effort="medium",
+            drop_params=True,
+        )
+
+    assert params["reasoning_effort"] == "medium"
+    assert not [record for record in caplog.records if "reasoning_effort" in record.getMessage()]

tests/unit/router_utils/test_reasoning_effort_capability.py

@@ -626,8 +626,9 @@ class TestResolveDeploymentReasoningEffort:
         _sent("turbo\nforged line", drop_params=True)
 
         assert [record.getMessage() for record in caplog.records] == [
-            "reasoning_effort unrecognized sent as None (rounding up) to deployment deployment-1, "
-            "model group kimi-k3, model hosted_vllm/my-kimi"
+            "Dropping reasoning level unrecognized (drop_params=True) for deployment deployment-1, model group "
+            "kimi-k3, model hosted_vllm/my-kimi because its model_info.reasoning_effort_levels does not list that "
+            "level. List every level the deployment supports."
         ]
 
     def test_line_breaks_in_logged_names_cannot_forge_a_log_line(self, caplog: pytest.LogCaptureFixture) -> None:
@@ -669,6 +670,27 @@ class TestResolveDeploymentReasoningEffort:
             "reasoning_effort_rounding": "up",
         }
 
+    def test_only_a_dropped_level_logs_a_warning(self, caplog: pytest.LogCaptureFixture) -> None:
+        verbose_logger.addHandler(caplog.handler)
+        verbose_logger.propagate = False
+        verbose_logger.setLevel(logging.INFO)
+        declared: Final = {"reasoning_effort_levels": ["low", "high"]}
+
+        _sent("none", declared, drop_params=True)
+        _sent("medium")
+        _sent("default", declared, drop_params=True)
+        _sent("default", declared)
+
+        assert [record.levelno for record in caplog.records] == [
+            logging.WARNING,
+            logging.INFO,
+            logging.INFO,
+            logging.INFO,
+        ]
+        dropped: Final = caplog.records[0].getMessage()
+        assert "Dropping reasoning level none" in dropped
+        assert "model_info.reasoning_effort_levels" in dropped
+
 
 class TestAzureGpt6SolAndLunaAdvertiseTheOpenAiLevels:
     @pytest.mark.parametrize(

tests/unit/test_utils_get_optional_params.py

@@ -1,4 +1,5 @@
 import json
+import logging
 from typing import Final
 from unittest.mock import MagicMock, patch
 
@@ -2097,3 +2098,127 @@ def test_supports_system_message_merges_system_prompt_into_user_message() -> Non
     ]
 
     assert map_system_message_pt(messages=messages) == [{"role": "user", "content": "Listen here! Hello there!"}]
+
+
+def test_dropping_an_unsupported_reasoning_effort_logs_one_warning(
+    local_model_cost_map: None, caplog: pytest.LogCaptureFixture
+) -> None:
+    with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+        optional_params: Final = get_optional_params(
+            model="terra-ptu", custom_llm_provider="azure", reasoning_effort="none", drop_params=True
+        )
+
+    assert "reasoning_effort" not in optional_params
+    drop_warnings: Final = [
+        record.getMessage() for record in caplog.records if "Dropping reasoning_effort" in record.getMessage()
+    ]
+    assert len(drop_warnings) == 1
+    assert all(fragment in drop_warnings[0] for fragment in ("'none'", "terra-ptu", "model_info.base_model"))
+
+
+def test_dropping_an_unsupported_thinking_logs_one_warning(
+    local_model_cost_map: None, caplog: pytest.LogCaptureFixture
+) -> None:
+    with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+        optional_params: Final = get_optional_params(
+            model="terra-ptu", custom_llm_provider="azure", thinking={"type": "disabled"}, drop_params=True
+        )
+
+    assert "thinking" not in optional_params
+    drop_warnings: Final = [
+        record.getMessage() for record in caplog.records if "Dropping thinking" in record.getMessage()
+    ]
+    assert len(drop_warnings) == 1
+    assert all(fragment in drop_warnings[0] for fragment in ("'disabled'", "terra-ptu", "model_info.base_model"))
+
+
+@pytest.mark.parametrize(
+    ("reasoning_kwargs", "warns"),
+    [
+        ({"reasoning_effort": "none"}, False),
+        ({"thinking": {"type": "disabled"}}, False),
+        ({"reasoning_effort": "high"}, True),
+        ({"thinking": {"type": "enabled", "budget_tokens": 1024}}, True),
+    ],
+)
+def test_reasoning_drop_warns_unless_a_mapped_non_reasoning_model_was_asked_for_none(
+    local_model_cost_map: None,
+    caplog: pytest.LogCaptureFixture,
+    reasoning_kwargs: dict[str, object],
+    warns: bool,
+) -> None:
+    litellm.register_model(
+        {
+            "azure/plain-chat-test-deployment": {
+                "litellm_provider": "azure",
+                "mode": "chat",
+                "max_tokens": 1024,
+                "input_cost_per_token": 0.0,
+                "output_cost_per_token": 0.0,
+                "supports_reasoning": False,
+            }
+        }
+    )
+
+    with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+        optional_params: Final = get_optional_params(
+            model="plain-chat-test-deployment", custom_llm_provider="azure", drop_params=True, **reasoning_kwargs
+        )
+
+    assert not set(reasoning_kwargs) & set(optional_params)
+    assert any("Dropping" in record.getMessage() for record in caplog.records) is warns
+
+
+@pytest.mark.parametrize("reasoning_kwargs", [{"reasoning_effort": "none"}, {"thinking": {"type": "disabled"}}])
+def test_turning_reasoning_off_warns_when_only_custom_pricing_describes_the_model(
+    local_model_cost_map: None, caplog: pytest.LogCaptureFixture, reasoning_kwargs: dict[str, object]
+) -> None:
+    litellm.register_model(
+        {
+            "azure/pricing-only-test-deployment": {
+                "litellm_provider": "azure",
+                "mode": "chat",
+                "input_cost_per_token": 0.000001,
+                "output_cost_per_token": 0.000002,
+            }
+        }
+    )
+
+    with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+        optional_params: Final = get_optional_params(
+            model="pricing-only-test-deployment", custom_llm_provider="azure", drop_params=True, **reasoning_kwargs
+        )
+
+    assert not set(reasoning_kwargs) & set(optional_params)
+    drop_warnings: Final = [record.getMessage() for record in caplog.records if "Dropping" in record.getMessage()]
+    assert len(drop_warnings) == 1
+    assert all(fragment in drop_warnings[0] for fragment in ("pricing-only-test-deployment", "model_info.base_model"))
+
+
+def test_dropping_reasoning_for_a_catalog_reasoning_model_does_not_suggest_base_model(
+    local_model_cost_map: None, caplog: pytest.LogCaptureFixture
+) -> None:
+    assert litellm.model_cost["azure/gpt-5.6-sol"].get("supports_reasoning") is True
+    with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+        optional_params: Final = get_optional_params(
+            model="gpt-5.6-sol", custom_llm_provider="azure", thinking={"type": "disabled"}, drop_params=True
+        )
+
+    assert "thinking" not in optional_params
+    drop_warnings: Final = [
+        record.getMessage() for record in caplog.records if "Dropping thinking" in record.getMessage()
+    ]
+    assert len(drop_warnings) == 1
+    assert "model_info.base_model" not in drop_warnings[0]
+
+
+def test_turning_reasoning_off_for_a_catalog_non_reasoning_model_drops_it_silently(
+    local_model_cost_map: None, caplog: pytest.LogCaptureFixture
+) -> None:
+    with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+        optional_params: Final = get_optional_params(
+            model="gpt-4o", custom_llm_provider="azure", reasoning_effort="none", drop_params=True
+        )
+
+    assert "reasoning_effort" not in optional_params
+    assert not [record for record in caplog.records if "Dropping" in record.getMessage()]