From df6091e59f3caab64f5eb02b0d0e072ca841e1bd Mon Sep 17 00:00:00 2001
From: Mateo Wang <277851410+mateo-berri@users.noreply.github.com>
Date: Thu, 3 Sep 2026 14:00:59 -0700
Subject: [PATCH 1/8] fix: treat gpt-6 names as the gpt-5 request family in
 OpenAI and Azure configs (#39631)

Adapted for stable/1.99.x: The Responses parametrize cases target a test class this line does not have, so the two gpt-6-astra cases are kept as a standalone test in the same file.

(cherry picked from commit 025a3ca42fa81017e961815ef229bd04ff9e903a)
---
 .../llms/azure/chat/gpt_5_transformation.py   | 22 ++--------------
 litellm/llms/azure/chat/gpt_transformation.py |  3 ++-
 .../llms/openai/chat/gpt_5_transformation.py  | 26 ++++++++-----------
 .../llms/openai/responses/transformation.py   |  3 ++-
 .../chat/test_azure_gpt5_transformation.py    | 12 +++++++++
 .../test_openai_responses_transformation.py   | 10 +++++++
 .../llms/openai/test_gpt5_transformation.py   | 14 ++++++++++
 .../llms/openai/test_is_model_gpt_5_model.py  |  4 +++
 tests/test_litellm/test_main.py               | 13 ++++++++++
 9 files changed, 70 insertions(+), 37 deletions(-)

diff --git a/litellm/llms/azure/chat/gpt_5_transformation.py b/litellm/llms/azure/chat/gpt_5_transformation.py
index d7584083327..d72f907fc88 100644
--- a/litellm/llms/azure/chat/gpt_5_transformation.py
+++ b/litellm/llms/azure/chat/gpt_5_transformation.py
@@ -7,6 +7,7 @@ from litellm.exceptions import UnsupportedParamsError
 from litellm.llms.openai.chat.gpt_5_transformation import (
     OpenAIGPT5Config,
     _get_effort_level,
+    is_gpt_reasoning_series_name,
 )
 from litellm.types.llms.openai import AllMessageValues
 
@@ -35,26 +36,7 @@ class AzureOpenAIGPT5Config(AzureOpenAIConfig, OpenAIGPT5Config):
 
     @classmethod
     def is_model_gpt_5_model(cls, model: str) -> bool:
-        """Check if the Azure model string refers to a gpt-5 variant.
-
-        Accepts both explicit gpt-5 model names and the ``gpt5_series/`` prefix
-        used for manual routing.
-        """
-        # The gpt-5-chat* family (gpt-5-chat, gpt-5-chat-latest, gpt-5-chat-2025-08-07,
-        # …) are regular chat models: they support temperature and tool_choice but NOT
-        # reasoning_effort.  They must NOT be routed through the GPT-5 reasoning path.
-        #
-        # Versioned chat models such as gpt-5.3-chat and gpt-5.1-chat ARE reasoning
-        # models and must stay on the GPT-5 path.  The distinguishing feature is that
-        # the gpt-5-chat family has a literal "-chat" immediately after "gpt-5"
-        # (i.e. "gpt-5-chat…"), while versioned chat models interpose a minor version
-        # number (i.e. "gpt-5.<digit>-chat").
-        #
-        # Using a startswith("gpt-5-chat") prefix check on the normalized name (rather
-        # than a substring check) makes this boundary explicit and avoids any ambiguity
-        # if future model names coincidentally contain "gpt-5-chat" as an interior run.
-        _normalized: Final = model.split("/")[-1]  # strip provider prefix, e.g. "azure/"
-        return ("gpt-5" in model and not _normalized.startswith("gpt-5-chat")) or "gpt5_series" in model
+        return is_gpt_reasoning_series_name(model) or "gpt5_series" in model
 
     def get_supported_openai_params(self, model: str) -> list[str]:
         """Get supported parameters for Azure OpenAI GPT-5 models.
diff --git a/litellm/llms/azure/chat/gpt_transformation.py b/litellm/llms/azure/chat/gpt_transformation.py
index 0d50609555a..940529f1add 100644
--- a/litellm/llms/azure/chat/gpt_transformation.py
+++ b/litellm/llms/azure/chat/gpt_transformation.py
@@ -10,6 +10,7 @@ from litellm.litellm_core_utils.prompt_templates.factory import (
     convert_to_azure_openai_messages,
 )
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
+from litellm.llms.openai.chat.gpt_5_transformation import GPT_REASONING_SERIES_MARKERS
 from litellm.types.llms.azure import (
     API_VERSION_MONTH_SUPPORTED_RESPONSE_FORMAT,
     API_VERSION_YEAR_SUPPORTED_RESPONSE_FORMAT,
@@ -120,7 +121,7 @@ class AzureOpenAIConfig(BaseConfig):
         name family needs the rename, including the ``gpt-5-chat*`` models that are excluded from
         the reasoning path by https://github.com/BerriAI/litellm/issues/13781.
         """
-        return "gpt-5" in model or "gpt5_series" in model
+        return any(marker in model for marker in GPT_REASONING_SERIES_MARKERS) or "gpt5_series" in model
 
     def _is_response_format_supported_model(self, model: str) -> bool:
         """
diff --git a/litellm/llms/openai/chat/gpt_5_transformation.py b/litellm/llms/openai/chat/gpt_5_transformation.py
index ffa3de0d5c6..acb56be47e1 100644
--- a/litellm/llms/openai/chat/gpt_5_transformation.py
+++ b/litellm/llms/openai/chat/gpt_5_transformation.py
@@ -44,6 +44,14 @@ def _get_effort_level(value: str | dict | None) -> str | None:
     return None
 
 
+GPT_REASONING_SERIES_MARKERS: Final = ("gpt-5", "gpt-6")
+
+
+def is_gpt_reasoning_series_name(model: str) -> bool:
+    normalized: Final = model.split("/")[-1]
+    return any(marker in model for marker in GPT_REASONING_SERIES_MARKERS) and not normalized.startswith("gpt-5-chat")
+
+
 class OpenAIGPT5Config(OpenAIGPTConfig):
     """Configuration for gpt-5 models including GPT-5-Codex variants.
 
@@ -56,21 +64,7 @@ class OpenAIGPT5Config(OpenAIGPTConfig):
 
     @classmethod
     def is_model_gpt_5_model(cls, model: str) -> bool:
-        # The gpt-5-chat* family (gpt-5-chat, gpt-5-chat-latest, gpt-5-chat-2025-08-07,
-        # …) are regular chat models: they support temperature and tool_choice but NOT
-        # reasoning_effort.  They must NOT be routed through the GPT-5 reasoning path.
-        #
-        # Versioned chat models such as gpt-5.3-chat and gpt-5.1-chat ARE reasoning
-        # models and must stay on the GPT-5 path.  The distinguishing feature is that
-        # the gpt-5-chat family has a literal "-chat" immediately after "gpt-5"
-        # (i.e. "gpt-5-chat…"), while versioned chat models interpose a minor version
-        # number (i.e. "gpt-5.<digit>-chat").
-        #
-        # Using a startswith("gpt-5-chat") prefix check on the normalized name (rather
-        # than a substring check) makes this boundary explicit and avoids any ambiguity
-        # if future model names coincidentally contain "gpt-5-chat" as an interior run.
-        _normalized: Final = model.split("/")[-1]  # strip provider prefix, e.g. "openai/"
-        return "gpt-5" in model and not _normalized.startswith("gpt-5-chat")
+        return is_gpt_reasoning_series_name(model)
 
     @classmethod
     def is_model_gpt_5_search_model(cls, model: str) -> bool:
@@ -105,6 +99,8 @@ class OpenAIGPT5Config(OpenAIGPTConfig):
     def is_model_gpt_5_4_plus_model(cls, model: str) -> bool:
         """Check if the model is gpt-5.4 or newer (5.4, 5.5, 5.6, etc., including pro)."""
         model_name: Final = model.split("/")[-1]
+        if model_name.startswith("gpt-6"):
+            return True
         if not model_name.startswith("gpt-5."):
             return False
         try:
diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index b2a69564908..6c3c988ef54 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -12,6 +12,7 @@ from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response impo
 )
 from litellm.litellm_core_utils.url_utils import encode_url_path_segment
 from litellm.llms.base_llm.responses.transformation import BaseResponsesAPIConfig
+from litellm.llms.openai.chat.gpt_5_transformation import is_gpt_reasoning_series_name
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.llms.openai import *
 from litellm.types.responses.main import *
@@ -48,7 +49,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
         parts: Final = model.split("/")
         if len(parts) > 1 and parts[0] not in ("openai",):
             return False
-        return "gpt-5" in model and "gpt-5-chat" not in model
+        return is_gpt_reasoning_series_name(model)
 
     @staticmethod
     def _supports_reasoning_effort_none(model: str) -> bool:
diff --git a/tests/test_litellm/llms/azure/chat/test_azure_gpt5_transformation.py b/tests/test_litellm/llms/azure/chat/test_azure_gpt5_transformation.py
index 83562331b9a..aaa3f3f1b1c 100644
--- a/tests/test_litellm/llms/azure/chat/test_azure_gpt5_transformation.py
+++ b/tests/test_litellm/llms/azure/chat/test_azure_gpt5_transformation.py
@@ -299,3 +299,15 @@ def test_azure_gpt5_1_does_not_support_logprobs(config: AzureOpenAIGPT5Config):
     supported_params = config.get_supported_openai_params(model="gpt-5.1")
     assert "logprobs" not in supported_params
     assert "top_logprobs" not in supported_params
+
+
+def test_azure_gpt_6_astra_takes_the_reasoning_series_request_shape():
+    params = litellm.get_optional_params(
+        model="gpt-6-astra",
+        custom_llm_provider="azure",
+        max_tokens=100,
+        reasoning_effort="max",
+    )
+    assert params["max_completion_tokens"] == 100
+    assert "max_tokens" not in params
+    assert params["reasoning_effort"] == "max"
diff --git a/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py b/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py
index c03c632363d..f89c9e5323b 100644
--- a/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py
+++ b/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py
@@ -1593,3 +1593,13 @@ class TestPromptCacheOptionsOnResponsesPath:
             "text": "hi",
             "prompt_cache_breakpoint": {"mode": "explicit"},
         }
+
+
+@pytest.mark.parametrize("effort", [None, "low"])
+def test_gpt_6_astra_drops_temperature_on_the_responses_path(effort):
+    mapped = OpenAIResponsesAPIConfig().map_openai_params(
+        response_api_optional_params={"temperature": 0, **({"reasoning": {"effort": effort}} if effort else {})},
+        model="gpt-6-astra",
+        drop_params=True,
+    )
+    assert "temperature" not in mapped
diff --git a/tests/test_litellm/llms/openai/test_gpt5_transformation.py b/tests/test_litellm/llms/openai/test_gpt5_transformation.py
index d279b119efe..d8e8feac1e3 100644
--- a/tests/test_litellm/llms/openai/test_gpt5_transformation.py
+++ b/tests/test_litellm/llms/openai/test_gpt5_transformation.py
@@ -1309,3 +1309,17 @@ def test_responses_gpt54_allow_temperature_effort_none(
         drop_params=False,
     )
     assert params["temperature"] == 0.7
+
+
+def test_gpt_6_astra_takes_the_reasoning_series_request_shape():
+    params = litellm.get_optional_params(
+        model="gpt-6-astra",
+        custom_llm_provider="openai",
+        max_tokens=100,
+        reasoning_effort="max",
+        verbosity="low",
+    )
+    assert params["max_completion_tokens"] == 100
+    assert "max_tokens" not in params
+    assert params["reasoning_effort"] == "max"
+    assert params["verbosity"] == "low"
diff --git a/tests/test_litellm/llms/openai/test_is_model_gpt_5_model.py b/tests/test_litellm/llms/openai/test_is_model_gpt_5_model.py
index 1095819c98c..107a1afb2c6 100644
--- a/tests/test_litellm/llms/openai/test_is_model_gpt_5_model.py
+++ b/tests/test_litellm/llms/openai/test_is_model_gpt_5_model.py
@@ -41,6 +41,8 @@ from litellm.llms.azure.chat.gpt_5_transformation import AzureOpenAIGPT5Config
 
 # Models that MUST be classified as GPT-5 (routed through GPT-5 reasoning path)
 GPT5_MODELS = [
+    "gpt-6-astra",
+    "openai/gpt-6-astra",
     "gpt-5",
     "gpt-5.1",
     "gpt-5.2",
@@ -120,6 +122,8 @@ class TestOpenAIGPT5ConfigIsModelGpt5Model:
 # /v1/responses bridge (when reasoning_effort is set and tools are passed) on
 # is_model_gpt_5_4_plus_model, so the gpt-5.6 family must land on the True side.
 GPT5_4_PLUS_MODELS = [
+    "gpt-6-astra",
+    "openai/gpt-6-astra",
     "gpt-5.4",
     "gpt-5.5",
     "gpt-5.5-pro",
diff --git a/tests/test_litellm/test_main.py b/tests/test_litellm/test_main.py
index 99b1cc826aa..55a843212ab 100644
--- a/tests/test_litellm/test_main.py
+++ b/tests/test_litellm/test_main.py
@@ -774,6 +774,19 @@ def test_responses_api_bridge_check_gpt_5_4_tools_plus_reasoning_routes_to_respo
     assert model_info.get("mode") == "responses"
 
 
+def test_responses_api_bridge_check_gpt_6_astra_tools_with_default_reasoning_routes_to_responses():
+    from litellm.main import responses_api_bridge_check
+
+    model_info, model = responses_api_bridge_check(
+        model="gpt-6-astra",
+        custom_llm_provider="openai",
+        tools=[{"type": "function", "function": {"name": "get_capital"}}],
+    )
+
+    assert model == "gpt-6-astra"
+    assert model_info.get("mode") == "responses"
+
+
 def test_responses_api_bridge_check_gpt_5_5_tools_plus_reasoning_routes_to_responses():
     """gpt-5.5+ with both tools and reasoning_effort should route to Responses API."""
     from litellm.main import responses_api_bridge_check

From 9cd7cde1f9a4661ddca95d3c8d0e027710fbde8f Mon Sep 17 00:00:00 2001
From: amasen02 <amasen02@users.noreply.github.com>
Date: Fri, 4 Sep 2026 15:37:53 +0530
Subject: [PATCH 2/8] fix(proxy): invalidate end-user spend counter and cache
 on budget reset (#39726)

Adapted for stable/1.99.x: this line predates budget rollover (#38514), so each end user's spend counter is zeroed through the existing counter_keys list instead of counter_resets, and its cache entry is evicted via end_user_cache_key.

Signed-off-by: amasen02 <amasen02@users.noreply.github.com>
(cherry picked from commit daced81f20a6b98c01beecf66fa997310d90bfb4)
---
 .../proxy/common_utils/reset_budget_job.py    | 15 +++++++++--
 .../common_utils/test_reset_budget_job.py     | 26 +++++++++++++++++++
 2 files changed, 39 insertions(+), 2 deletions(-)

diff --git a/litellm/proxy/common_utils/reset_budget_job.py b/litellm/proxy/common_utils/reset_budget_job.py
index 8fcb184b26a..d8ff61d713d 100644
--- a/litellm/proxy/common_utils/reset_budget_job.py
+++ b/litellm/proxy/common_utils/reset_budget_job.py
@@ -32,7 +32,7 @@ from litellm.proxy.common_utils.timezone_utils import (
     compute_budget_reset_at,
     get_budget_reset_settings,
 )
-from litellm.proxy.common_utils.user_api_key_cache import tag_cache_key
+from litellm.proxy.common_utils.user_api_key_cache import end_user_cache_key, tag_cache_key
 from litellm.proxy.db.db_transaction_queue.pod_lock_manager import PodLockManager
 from litellm.proxy.db.exception_handler import call_with_db_reconnect_retry
 from litellm.proxy.utils import PrismaClient, ProxyLogging
@@ -122,6 +122,14 @@ def _tag_cache_keys(row: _TagRow) -> tuple[str, ...]:
     return (tag_cache_key(row.tag_name),)
 
 
+def _enduser_counter_key(row: _EndUserRow) -> str:
+    return f"spend:end_user:{row.user_id}"
+
+
+def _enduser_cache_keys(row: _EndUserRow) -> tuple[str, ...]:
+    return (end_user_cache_key(row.user_id),)
+
+
 def _budget_link_where(
     budget_ids: Sequence[str],
     extra: Mapping[str, object] = MappingProxyType({}),
@@ -522,6 +530,7 @@ class ResetBudgetJob:
             where=_budget_link_where(budget_ids, _SPENT_ROWS_WHERE),
             log_subject="tags",
         )
+        endusers: Final[tuple[_EndUserRow, ...]] = await self._collect_endusers_to_reset(budget_ids)
         return _BudgetCascade(
             budgets=tuple(budgets_to_reset),
             budget_ids=budget_ids,
@@ -533,18 +542,20 @@ class ResetBudgetJob:
                 for b in budgets_to_reset
                 if b.budget_id is not None and b.budget_duration is not None
             ),
-            endusers=await self._collect_endusers_to_reset(budget_ids),
+            endusers=endusers,
             counter_keys=(
                 *(_team_membership_counter_key(row) for row in team_memberships),
                 *(_key_counter_key(row) for row in keys),
                 *(_org_counter_key(row) for row in orgs),
                 *(_tag_counter_key(row) for row in tags),
+                *(_enduser_counter_key(row) for row in endusers),
             ),
             cache_keys=(
                 *(key for row in team_memberships for key in _team_membership_cache_keys(row)),
                 *(key for row in keys for key in _key_cache_keys(row)),
                 *(key for row in orgs for key in _org_cache_keys(row)),
                 *(key for row in tags for key in _tag_cache_keys(row)),
+                *(key for row in endusers for key in _enduser_cache_keys(row)),
             ),
         )
 
diff --git a/tests/test_litellm/proxy/common_utils/test_reset_budget_job.py b/tests/test_litellm/proxy/common_utils/test_reset_budget_job.py
index 25c177a308d..a6d0c3b5fa9 100644
--- a/tests/test_litellm/proxy/common_utils/test_reset_budget_job.py
+++ b/tests/test_litellm/proxy/common_utils/test_reset_budget_job.py
@@ -1346,6 +1346,32 @@ def test_budget_table_reset_invalidates_every_tag_not_just_the_first(reset_budge
     assert deleted == {"tag:tenant-a", "tag:tenant-b", "tag:tenant-c"}
 
 
+def test_budget_table_reset_invalidates_enduser_counter_and_cache(reset_budget_job, mock_prisma_client, monkeypatch):
+    """When an end user's budget resets, its Redis spend counter is zeroed and its management cache is evicted."""
+    counter_cache = _make_counter_invalidation_job(monkeypatch)
+    budget = _budget_row(budget_id="budget-1")
+    mock_prisma_client.data["budget"] = [budget]
+    test_enduser = type(
+        "LiteLLM_EndUserTable",
+        (),
+        {
+            "spend": 20.0,
+            "litellm_budget_table": budget,
+            "budget_id": "budget-1",
+            "user_id": "customer-42",
+        },
+    )
+    mock_prisma_client.data["enduser"] = [test_enduser]
+
+    asyncio.run(reset_budget_job.reset_budget_for_litellm_budget_table())
+
+    counter_cache.in_memory_cache.set_cache.assert_any_call(key="spend:end_user:customer-42", value=0.0, ttl=60)
+    counter_cache.redis_cache.async_set_cache.assert_any_await(key="spend:end_user:customer-42", value=0.0, ttl=60)
+    deleted = {call.kwargs.get("key") for call in counter_cache.user_api_key_cache.async_delete_cache.await_args_list}
+    assert "end_user_id:customer-42" in deleted
+
+
+
 def test_budget_table_reset_commits_even_when_cache_eviction_fails(reset_budget_job, mock_prisma_client, monkeypatch):
     """Eviction runs after the commit, so a broken cache cannot undo the write."""
     counter_cache = _make_counter_invalidation_job(monkeypatch)

From 1c6e6b1a731d395f69415208ca946ace2793dd18 Mon Sep 17 00:00:00 2001
From: ryan-crabbe-berri <ryan@berri.ai>
Date: Thu, 10 Sep 2026 17:31:25 -0700
Subject: [PATCH 3/8] fix(reset_budget_job): reset end users by budget link,
 not by user id (#40639)

Adapted for stable/1.99.x: this line predates budget rollover (#38514), so the fix is applied to _commit_budget_cascade_once directly. End users reset on the budget link plus a NULL budget_id branch for the default tier, which is what upstream's _queue_enduser_resets does with rollover off. The rollover test hunk is dropped.

(cherry picked from commit 8a4fae0e174799ce2d3858a55473be0d6247a2a4)
---
 .../proxy/common_utils/reset_budget_job.py    |  6 +-
 .../test_proxy_budget_reset.py                |  6 +-
 .../common_utils/test_reset_budget_job.py     | 73 ++++++++++++++++---
 3 files changed, 68 insertions(+), 17 deletions(-)

diff --git a/litellm/proxy/common_utils/reset_budget_job.py b/litellm/proxy/common_utils/reset_budget_job.py
index d8ff61d713d..2cb6bafbc21 100644
--- a/litellm/proxy/common_utils/reset_budget_job.py
+++ b/litellm/proxy/common_utils/reset_budget_job.py
@@ -576,14 +576,14 @@ class ResetBudgetJob:
         )
 
     async def _commit_budget_cascade_once(self, cascade: _BudgetCascade) -> None:
-        enduser_ids: Final = tuple(row.user_id for row in cascade.endusers)
         async with budget_cascade_unit_of_work(self.prisma_client.db.batch_) as uow:
             uow.team_memberships.queue_spend_zero(where=_budget_link_where(cascade.budget_ids))
             uow.keys.queue_spend_zero(where=_budget_link_where(cascade.budget_ids, _LINKED_KEYS_WHERE))
             uow.organizations.queue_spend_zero(where=_budget_link_where(cascade.budget_ids, _SPENT_ROWS_WHERE))
             uow.tags.queue_spend_zero(where=_budget_link_where(cascade.budget_ids, _SPENT_ROWS_WHERE))
-            if enduser_ids:
-                uow.endusers.queue_spend_zero(where={"user_id": {"in": list(enduser_ids)}})
+            uow.endusers.queue_spend_zero(where=_budget_link_where(cascade.budget_ids, _SPENT_ROWS_WHERE))
+            if litellm.max_end_user_budget_id in cascade.budget_ids:
+                uow.endusers.queue_spend_zero(where={"budget_id": None, **_SPENT_ROWS_WHERE})
             for budget_id, budget_reset_at in cascade.budget_resets:
                 uow.budgets.queue_window_advance(budget_id=budget_id, budget_reset_at=budget_reset_at)
 
diff --git a/tests/litellm_utils_tests/test_proxy_budget_reset.py b/tests/litellm_utils_tests/test_proxy_budget_reset.py
index 9f6e1f4c3f7..4103536950d 100644
--- a/tests/litellm_utils_tests/test_proxy_budget_reset.py
+++ b/tests/litellm_utils_tests/test_proxy_budget_reset.py
@@ -401,7 +401,7 @@ async def test_reset_budget_endusers_are_zeroed_with_the_budget_window_advance()
 
     enduser_writes = [c for c in batch_calls if c["table"] == "enduser"]
     assert len(enduser_writes) == 1
-    assert enduser_writes[0]["where"]["user_id"]["in"] == [f"user{i}" for i in range(1, 7)]
+    assert enduser_writes[0]["where"] == {"budget_id": {"in": ["budget1"]}, "spend": {"gt": 0}}
     assert enduser_writes[0]["data"] == {"spend": 0}
 
     budget_writes = [c for c in batch_calls if c["table"] == "budget"]
@@ -602,7 +602,7 @@ async def test_reset_budget_continues_other_categories_on_failure():
     assert len([c for c in batch_calls if c["table"] == "team_membership"]) == 1
     enduser_writes = [c for c in batch_calls if c["table"] == "enduser"]
     assert len(enduser_writes) == 1
-    assert enduser_writes[0]["where"] == {"user_id": {"in": ["user1"]}}
+    assert enduser_writes[0]["where"] == {"budget_id": {"in": ["budget1"]}, "spend": {"gt": 0}}
     assert enduser_writes[0]["data"] == {"spend": 0}
 
     # Check the new batch write path: 2 keys + 1 user (user1 failed) + 2 teams.
@@ -1031,7 +1031,7 @@ async def test_service_logger_endusers_success():
 
     enduser_writes = [c for c in batch_calls if c["table"] == "enduser"]
     assert len(enduser_writes) == 1
-    assert enduser_writes[0]["where"] == {"user_id": {"in": ["user1", "user2"]}}
+    assert enduser_writes[0]["where"] == {"budget_id": {"in": ["budget1"]}, "spend": {"gt": 0}}
 
     proxy_logging_obj.service_logging_obj.async_service_success_hook.assert_called_once()
     (
diff --git a/tests/test_litellm/proxy/common_utils/test_reset_budget_job.py b/tests/test_litellm/proxy/common_utils/test_reset_budget_job.py
index a6d0c3b5fa9..3822e78516b 100644
--- a/tests/test_litellm/proxy/common_utils/test_reset_budget_job.py
+++ b/tests/test_litellm/proxy/common_utils/test_reset_budget_job.py
@@ -4,7 +4,7 @@ import sys
 import types
 from datetime import datetime, timedelta, timezone
 from datetime import time as dt_time
-from typing import Any, Dict, List
+from typing import Any, Dict, Final, List
 from unittest.mock import AsyncMock, MagicMock
 
 import httpx
@@ -401,7 +401,7 @@ def test_reset_budget_for_enduser(reset_budget_job, mock_prisma_client):
         {
             "table": "enduser",
             "op": "update_many",
-            "where": {"user_id": {"in": ["test-enduser-1"]}},
+            "where": {"budget_id": {"in": ["test-budget-1"]}, "spend": {"gt": 0}},
             "data": {"spend": 0},
         }
     ]
@@ -502,7 +502,7 @@ def test_reset_budget_all(reset_budget_job, mock_prisma_client):
         {
             "table": "enduser",
             "op": "update_many",
-            "where": {"user_id": {"in": ["test-enduser-1"]}},
+            "where": {"budget_id": {"in": ["test-budget-1"]}, "spend": {"gt": 0}},
             "data": {"spend": 0},
         }
     ]
@@ -521,6 +521,7 @@ _LINKED_TABLE_CASES = [
     ),
     ("org", {"budget_id": {"in": ["7d-budget-tier"]}, "spend": {"gt": 0}}),
     ("tag", {"budget_id": {"in": ["7d-budget-tier"]}, "spend": {"gt": 0}}),
+    ("enduser", {"budget_id": {"in": ["7d-budget-tier"]}, "spend": {"gt": 0}}),
 ]
 
 
@@ -550,6 +551,48 @@ def test_budget_table_reset_zeroes_spend_on_every_linked_table(
     assert writes[0]["data"] == {"spend": 0}
 
 
+_POSTGRES_MAX_BIND_VARIABLES: Final = 32767
+
+
+def _bind_count(where: Dict[str, Any]) -> int:
+    """Bind variables one prisma where-clause compiles to: each scalar is one
+    placeholder and an ``in`` list contributes one per element."""
+    return sum(len(value["in"]) if isinstance(value, dict) and "in" in value else 1 for value in where.values())
+
+
+@pytest.mark.parametrize("population", [3, 40_000], ids=["small", "over-pg-bind-ceiling"])
+def test_enduser_reset_bind_count_does_not_scale_with_population(reset_budget_job, mock_prisma_client, population):
+    """Regression for #40564.
+
+    Enumerating every dependent user id put one bind variable per customer into
+    a single prepared statement. Past PostgreSQL's ceiling the statement could
+    not be parsed at all, so the whole atomic cascade rolled back,
+    budget_reset_at never advanced, and the tier stayed due on every later tick
+    forever. Matching on the budget link keeps the statement the same size no
+    matter how many customers share a tier.
+    """
+    budget = _budget_row(budget_id="shared-tier", budget_duration="1d")
+    mock_prisma_client.data["budget"] = [budget]
+    mock_prisma_client.data["enduser"] = [
+        types.SimpleNamespace(
+            spend=1.0,
+            litellm_budget_table=budget,
+            user_id=f"cust-{index:08d}",
+            budget_id="shared-tier",
+        )
+        for index in range(population)
+    ]
+
+    asyncio.run(reset_budget_job.reset_budget_for_litellm_budget_table())
+
+    writes = _batch_writes(mock_prisma_client, "enduser")
+    assert [_bind_count(write["where"]) for write in writes] == [2], (
+        f"the cascade must not enumerate {population} user ids: past "
+        f"{_POSTGRES_MAX_BIND_VARIABLES} binds PostgreSQL refuses the statement, got {writes[:1]}"
+    )
+    assert _batch_writes(mock_prisma_client, "budget")[0]["data"]["budget_reset_at"] is not None
+
+
 def test_budget_table_reset_writes_nothing_when_no_budget_is_due(reset_budget_job, mock_prisma_client):
     """Nothing due means no transaction is opened at all."""
     asyncio.run(reset_budget_job.reset_budget_for_litellm_budget_table())
@@ -717,14 +760,22 @@ def test_reset_budget_resets_endusers_with_null_budget_id(reset_budget_job, mock
 
     asyncio.run(reset_budget_job.reset_budget_for_litellm_budget_table())
 
-    # Both end users are zeroed by the same committed statement.
-    enduser_writes = _batch_writes(mock_prisma_client, "enduser")
-    assert len(enduser_writes) == 1, f"Expected a single enduser write, got {enduser_writes}"
-    assert set(enduser_writes[0]["where"]["user_id"]["in"]) == {
-        "enduser-explicit",
-        "enduser-implicit",
-    }
-    assert enduser_writes[0]["data"] == {"spend": 0}
+    # Both end users are zeroed: the linked rows on the tier's budget_id, the
+    # implicit ones on the NULL branch that stands in for the default tier.
+    assert _batch_writes(mock_prisma_client, "enduser") == [
+        {
+            "table": "enduser",
+            "op": "update_many",
+            "where": {"budget_id": {"in": [default_budget_id]}, "spend": {"gt": 0}},
+            "data": {"spend": 0},
+        },
+        {
+            "table": "enduser",
+            "op": "update_many",
+            "where": {"budget_id": None, "spend": {"gt": 0}},
+            "data": {"spend": 0},
+        },
+    ]
 
     # Verify find_many was called to fetch NULL-budget-id end users
     find_many_calls = mock_prisma_client.db.litellm_endusertable.find_many_calls

From a790967d62627eb76ca6e1fe73cf98059e63ce10 Mon Sep 17 00:00:00 2001
From: Yuneng Jiang <yuneng@berri.ai>
Date: Thu, 24 Sep 2026 14:47:07 -0700
Subject: [PATCH 4/8] chore(deps): bump anyio to 4.14.2

---
 uv.lock | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/uv.lock b/uv.lock
index 82f06a3a3c2..c39c8ae4222 100644
--- a/uv.lock
+++ b/uv.lock
@@ -10,7 +10,7 @@ resolution-markers = [
 ]
 
 [options]
-exclude-newer = "2026-09-12T23:03:29.512701Z"
+exclude-newer = "2026-09-21T21:47:06.53676Z"
 exclude-newer-span = "P3D"
 
 [manifest]
@@ -315,16 +315,16 @@ vertex = [
 
 [[package]]
 name = "anyio"
-version = "4.13.0"
+version = "4.14.2"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "exceptiongroup", marker = "python_full_version < '3.11'" },
     { name = "idna" },
     { name = "typing-extensions", marker = "python_full_version < '3.13'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/19/14/2c5dd9f512b66549ae92767a9c7b330ae88e1932ca57876909410251fe13/anyio-4.13.0.tar.gz", hash = "sha256:334b70e641fd2221c1505b3890c69882fe4a2df910cba14d97019b90b24439dc", size = 231622, upload-time = "2026-03-24T12:59:09.671Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/61/cc/a381afa6efea9f496eff839d4a6a1aed3bfafc7b3ab4b0d1b243a12573dd/anyio-4.14.2.tar.gz", hash = "sha256:cfa139f3ed1a23ee8f88a145ddb5ac7605b8bbfd8592baacd7ce3d8bb4313c7f", size = 260176, upload-time = "2026-07-12T20:29:07.082Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/da/42/e921fccf5015463e32a3cf6ee7f980a6ed0f395ceeaa45060b61d86486c2/anyio-4.13.0-py3-none-any.whl", hash = "sha256:08b310f9e24a9594186fd75b4f73f4a4152069e3853f1ed8bfbf58369f4ad708", size = 114353, upload-time = "2026-03-24T12:59:08.246Z" },
+    { url = "https://files.pythonhosted.org/packages/da/35/f2287558c17e29fafc8ef3daf819bb9834061cfa43bff8014f7df7f63bdc/anyio-4.14.2-py3-none-any.whl", hash = "sha256:9f505dda5ac9f0c8309b5e8bd445a8c2bf7246f3ce950121e45ea15bc41d1494", size = 125813, upload-time = "2026-07-12T20:29:05.763Z" },
 ]
 
 [[package]]

From fd8adafa50a91e6b7355b9a4ac29d6c1c70a74b0 Mon Sep 17 00:00:00 2001
From: Yuneng Jiang <yuneng@berri.ai>
Date: Thu, 24 Sep 2026 14:47:08 -0700
Subject: [PATCH 5/8] chore(deps): bump gitpython to 3.1.60

---
 uv.lock | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/uv.lock b/uv.lock
index c39c8ae4222..14028400aec 100644
--- a/uv.lock
+++ b/uv.lock
@@ -10,7 +10,7 @@ resolution-markers = [
 ]
 
 [options]
-exclude-newer = "2026-09-21T21:47:06.53676Z"
+exclude-newer = "2026-09-21T21:47:07.908357Z"
 exclude-newer-span = "P3D"
 
 [manifest]
@@ -2373,14 +2373,14 @@ wheels = [
 
 [[package]]
 name = "gitpython"
-version = "3.1.59"
+version = "3.1.60"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "gitdb" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/ca/dc/126b28e76b24a9268ba931ad3e012f71ebdadf62fd9f17758f7074bb0b20/gitpython-3.1.59.tar.gz", hash = "sha256:0a1475cfdc38a5bfba1a3e9a4a9da52a39749ecec322b772915c019f94e5b7e4", size = 230445, upload-time = "2026-08-10T12:03:20.271Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/84/14/e6b1a48d831755a53c2029351fcef82e70db4a08f338daefe29d8d0cf31c/gitpython-3.1.60.tar.gz", hash = "sha256:e936431879fa85581b4311fa63492ea52251909e2d655b6529c704c904ddcc24", size = 230793, upload-time = "2026-08-25T18:33:46.102Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/ef/ed/ae57eb7d344f43f87b74b3a281ead6ec7d6394eef72a7b1dcb28dd089550/gitpython-3.1.59-py3-none-any.whl", hash = "sha256:67a82f537384578643624c8b2c531938a9b82be431663e575dcf638526631d4c", size = 220996, upload-time = "2026-08-10T12:03:18.804Z" },
+    { url = "https://files.pythonhosted.org/packages/71/63/ba28697918b7c190af9f3f21940d03e8814e25dd4ddd39d6929f3a553995/gitpython-3.1.60-py3-none-any.whl", hash = "sha256:39548bffb8fa0f3a548133348868bb4838e79d73283052207dc97781a569b6b4", size = 221893, upload-time = "2026-08-25T18:33:44.75Z" },
 ]
 
 [[package]]

From d65a650250d406ddf16e649f1e858fbc0f7e8aee Mon Sep 17 00:00:00 2001
From: Yuneng Jiang <yuneng@berri.ai>
Date: Thu, 24 Sep 2026 14:47:09 -0700
Subject: [PATCH 6/8] chore(deps): bump soupsieve to 2.9

---
 uv.lock | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/uv.lock b/uv.lock
index 14028400aec..160dcb242eb 100644
--- a/uv.lock
+++ b/uv.lock
@@ -10,7 +10,7 @@ resolution-markers = [
 ]
 
 [options]
-exclude-newer = "2026-09-21T21:47:07.908357Z"
+exclude-newer = "2026-09-21T21:47:08.546881Z"
 exclude-newer-span = "P3D"
 
 [manifest]
@@ -8955,11 +8955,11 @@ wheels = [
 
 [[package]]
 name = "soupsieve"
-version = "2.8.4"
+version = "2.9"
 source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/47/2c/0a5f6f8ee0d5589e48c7640213ed5175d52cf540a06725b628cc1a45d6ce/soupsieve-2.8.4.tar.gz", hash = "sha256:e121fd02e975c695e4e9e8774a5ee35d74714b59307868dcc5319ad2d9e3328e", size = 121110, upload-time = "2026-05-24T13:55:57.154Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/80/f1/93422647dd7e461f23d254e6b2bfa687a85b53aeb4903fcdbb74474d4584/soupsieve-2.9.tar.gz", hash = "sha256:acee8417325c5653e1377dc31eccad59eb82cbc65942afe6174c53b3aaad63fc", size = 122122, upload-time = "2026-07-19T01:35:18.425Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/5e/f5/0c41cb68dcae6b7de4fac4188a3a9589e21fb31df21ea3a2e888db95e6c9/soupsieve-2.8.4-py3-none-any.whl", hash = "sha256:e7e6b0769c8f51ed59acab6e994b00621096cfb1c640a7509295987388fbaf65", size = 37304, upload-time = "2026-05-24T13:55:55.406Z" },
+    { url = "https://files.pythonhosted.org/packages/7b/d6/3185ab5ad1280319b31986898f3206dd7227cd75e293d4dba2a5e6bf27a0/soupsieve-2.9-py3-none-any.whl", hash = "sha256:a2b2c76d67df2382d245409fd71e321a571717e58463efa32ace87dcadac2c12", size = 37387, upload-time = "2026-07-19T01:35:17.106Z" },
 ]
 
 [[package]]

From 889ea16af7b038463ce692d6436ae22599582670 Mon Sep 17 00:00:00 2001
From: Yuneng Jiang <yuneng@berri.ai>
Date: Thu, 24 Sep 2026 20:39:37 -0700
Subject: [PATCH 7/8] =?UTF-8?q?bump:=20version=201.99.3=20=E2=86=92=201.99?=
 =?UTF-8?q?.4?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 pyproject.toml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/pyproject.toml b/pyproject.toml
index ae4a79e98a7..566086a177b 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "litellm"
-version = "1.99.3"
+version = "1.99.4"
 description = "Library to easily interface with LLM API providers"
 readme = "README.md"
 requires-python = ">=3.10, <3.15"
@@ -310,7 +310,7 @@ members = ["enterprise", "litellm-proxy-extras"]
 profile = "black"
 
 [tool.commitizen]
-version = "1.99.3"
+version = "1.99.4"
 version_files = [
     "pyproject.toml:^version",
 ]

From adf3ca0d04919da2e60e3172c948b44d8fd732ec Mon Sep 17 00:00:00 2001
From: Yuneng Jiang <yuneng@berri.ai>
Date: Thu, 24 Sep 2026 20:39:37 -0700
Subject: [PATCH 8/8] chore: refresh uv.lock for 1.99.4

---
 uv.lock | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/uv.lock b/uv.lock
index 160dcb242eb..7a919205fb9 100644
--- a/uv.lock
+++ b/uv.lock
@@ -10,7 +10,7 @@ resolution-markers = [
 ]
 
 [options]
-exclude-newer = "2026-09-21T21:47:08.546881Z"
+exclude-newer = "2026-09-22T03:39:37.254485Z"
 exclude-newer-span = "P3D"
 
 [manifest]
@@ -4266,7 +4266,7 @@ wheels = [
 
 [[package]]
 name = "litellm"
-version = "1.99.3"
+version = "1.99.4"
 source = { editable = "." }
 dependencies = [
     { name = "aiohttp" },