From cf98a333748736f2480f0fa76ff7a8706136f17d Mon Sep 17 00:00:00 2001
From: aayushbaluni <73417844+aayushbaluni@users.noreply.github.com>
Date: Tue, 18 Aug 2026 19:55:58 +0530
Subject: [PATCH] docs: state the mixed-group precedence for strict_token_count

A deployment carrying strict_token_count: false does not short-circuit to permissive; it falls through to the router check, so a group is strict if any deployment asks for it. Greptile flagged this as non-obvious. Documented on the helper and pinned with a test, so it is a contract rather than an accident.
---
 litellm/proxy/proxy_server.py                 |  8 +++
 .../test_strict_token_count.py                | 51 +++++++++++++++++++
 2 files changed, 59 insertions(+)

diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py
index 88f99664be8..b74c0872ce2 100644
--- a/litellm/proxy/proxy_server.py
+++ b/litellm/proxy/proxy_server.py
@@ -11704,6 +11704,14 @@ def _is_strict_token_count_model(
     router's configuration for the requested model. The fallback matters
     because deployment selection can fail for reasons unrelated to the
     policy, and a strict model must not quietly return an estimate then.
+
+    Note the precedence when a model group is configured inconsistently: a
+    deployment carrying `strict_token_count: false` does **not** short-circuit
+    to permissive. It falls through to the router check, so the group is
+    strict if any of its deployments asks for it. The caller cannot choose
+    which deployment serves them, so an ambiguous configuration is read the
+    safe way. Set the flag on every deployment in a group to avoid relying on
+    this.
     """
     if model_info is not None and bool(model_info.get("strict_token_count", False)):
         return True
diff --git a/tests/proxy_unit_tests/test_strict_token_count.py b/tests/proxy_unit_tests/test_strict_token_count.py
index 66cfc3d67ae..6d30f907582 100644
--- a/tests/proxy_unit_tests/test_strict_token_count.py
+++ b/tests/proxy_unit_tests/test_strict_token_count.py
@@ -235,6 +235,57 @@ async def test_non_strict_model_still_estimates_when_selection_fails():
     assert response.total_tokens > 0
 
 
+@pytest.mark.asyncio
+async def test_mixed_group_is_strict_if_any_deployment_asks():
+    """A group configured inconsistently resolves to strict.
+
+    A deployment carrying `strict_token_count: false` does not make the group
+    permissive when a sibling asks for strict. The caller cannot choose which
+    deployment serves them, so the ambiguous configuration is read the safe
+    way. Documented on `_is_strict_token_count_model` because it is
+    non-obvious. See https://github.com/BerriAI/litellm/issues/37102.
+    """
+    router = Router(
+        model_list=[
+            {
+                "model_name": "mixed-model",
+                "litellm_params": {
+                    "model": "bedrock/anthropic.claude-opus-5-20260101-v1:0",
+                    "aws_region_name": "us-east-1",
+                    "aws_access_key_id": "fake",
+                    "aws_secret_access_key": "fake",
+                },
+                "model_info": {"strict_token_count": False},
+            },
+            {
+                "model_name": "mixed-model",
+                "litellm_params": {
+                    "model": "bedrock/anthropic.claude-opus-5-20260101-v1:0",
+                    "aws_region_name": "us-west-2",
+                    "aws_access_key_id": "fake",
+                    "aws_secret_access_key": "fake",
+                },
+                "model_info": {"strict_token_count": True},
+            },
+        ]
+    )
+
+    original_router = getattr(proxy_server, "llm_router", None)
+    setattr(proxy_server, "llm_router", router)
+    try:
+        with _unsupported_count_tokens():
+            with pytest.raises(ProxyException):
+                await token_counter(
+                    request=TokenCountRequest(
+                        model="mixed-model",
+                        messages=[{"role": "user", "content": "hello " * 400}],
+                    ),
+                    call_endpoint=True,
+                )
+    finally:
+        setattr(proxy_server, "llm_router", original_router)
+
+
 @pytest.mark.asyncio
 async def test_disable_token_counter_still_applies_proxy_wide():
     """The existing proxy-wide flag must keep working for unmarked models."""