fix(proxy): always emit the Anthropic /v1/models token limits, null when unknown (#36961)

Anthropic's Models API declares max_input_tokens and max_tokens as nullable, not
optional, and the live vendor endpoint returns both keys on every entry. The
merged Anthropic-native listing dropped either key whenever LiteLLM could not
resolve a limit, so a client validating against a nullable-but-required schema
saw a malformed entry for any model the cost map does not know.

litellm/llms/anthropic/common_utils.py

@@ -1227,16 +1227,13 @@ def process_anthropic_headers(headers: httpx.Headers | dict) -> dict:
 
 
 def _anthropic_model_entry(model: ModelInfoResponse, created_at: str) -> Mapping[str, object]:
-    token_limits: Final = (
-        ("max_input_tokens", model.get("max_input_tokens")),
-        ("max_tokens", model.get("max_output_tokens")),
-    )
     return {  # mutable-ok: JSON response body, serialized by the route and never mutated
         "type": "model",
         "id": model["id"],
         "display_name": model["id"],
         "created_at": created_at,
-        **{name: limit for name, limit in token_limits if limit is not None},  # mutable-ok: merged into the body above
+        "max_input_tokens": model.get("max_input_tokens"),
+        "max_tokens": model.get("max_output_tokens"),
     }
 
 
@@ -1246,7 +1243,8 @@ def create_anthropic_model_list_response(models: Sequence[ModelInfoResponse]) ->
     Clients that send an anthropic-version header parse the Anthropic Models API
     shape (type/display_name/created_at plus has_more/first_id/last_id) and filter
     the list themselves, so every model is returned here. The token limits carry
-    over from the OpenAI-shaped listing, named as the Messages API names them
+    over from the OpenAI-shaped listing, named as the Messages API names them, and
+    are always present because the vendor shape declares them nullable, not optional
     """
     created_at: Final = (
         datetime.fromtimestamp(DEFAULT_MODEL_CREATED_AT_TIME, tz=timezone.utc).isoformat().replace("+00:00", "Z")

tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py

@@ -2056,11 +2056,14 @@ def test_create_anthropic_model_list_response_shape():
         # ISO 8601 with a Z suffix, as the Anthropic Models API returns.
         assert entry["created_at"].endswith("Z")
         assert "+00:00" not in entry["created_at"]
-        assert "max_input_tokens" not in entry
-        assert "max_tokens" not in entry
+        assert entry["max_input_tokens"] is None
+        assert entry["max_tokens"] is None
 
 
 def test_create_anthropic_model_list_response_carries_token_limits():
+    """max_input_tokens and max_tokens are nullable in the Anthropic Models shape,
+    not optional, so both keys are emitted for every entry and carry null when the
+    limit is unknown."""
     from litellm.llms.anthropic.common_utils import (
         create_anthropic_model_list_response,
     )
@@ -2091,9 +2094,12 @@ def test_create_anthropic_model_list_response_carries_token_limits():
     assert opus["max_tokens"] == 64000
     assert "max_output_tokens" not in opus
     assert input_only["max_input_tokens"] == 8192
-    assert "max_tokens" not in input_only
-    assert "max_input_tokens" not in unknown
-    assert "max_tokens" not in unknown
+    assert input_only["max_tokens"] is None
+    assert unknown["max_input_tokens"] is None
+    assert unknown["max_tokens"] is None
+    for entry in response["data"]:
+        assert "max_input_tokens" in entry
+        assert "max_tokens" in entry
 
 
 def test_create_anthropic_model_list_response_empty():

tests/test_litellm/proxy/proxy_server/test_routes_models.py

@@ -15,6 +15,8 @@ import pytest
 
 import litellm
 from litellm.proxy import proxy_server
+from litellm.proxy import utils as proxy_utils
+from litellm.proxy.utils import create_model_info_response
 
 from .conftest import normalize  # type: ignore[import-not-found]
 
@@ -151,8 +153,38 @@ def test_anthropic_format_exposes_token_limits(
     assert claude["max_input_tokens"] == 200000
     assert claude["max_tokens"] == 64000
     assert "max_output_tokens" not in claude
-    assert "max_input_tokens" not in gpt_4
-    assert "max_tokens" not in gpt_4
+    assert gpt_4["max_input_tokens"] is None
+    assert gpt_4["max_tokens"] is None
+
+
+@pytest.mark.parametrize("path", ["/v1/models", "/models"])
+def test_anthropic_format_carries_router_configured_token_limits(client, auth_as, patched_models, monkeypatch, path):
+    """Pins the whole resolution chain, not just the formatter: a deployment's
+    configured limits beat the cost map, and the configured output budget is what
+    lands on the Anthropic ``max_tokens``. All eight limits differ, so an entry
+    built from another entry's lookup shows up as the wrong numbers."""
+
+    def _configured(model_name):
+        return (300000, 32000) if model_name == "gpt-4" else (500000, 4096)
+
+    def _cost_map_lookup(model_id):
+        max_input, max_output = (200000, 64000) if model_id == "gpt-4" else (100000, 8000)
+        return {"max_input_tokens": max_input, "max_output_tokens": max_output, "mode": "chat"}
+
+    patched_models.get_configured_token_limits = MagicMock(side_effect=_configured)
+
+    def _resolved(**kwargs):
+        return create_model_info_response(**kwargs, get_model_info=_cost_map_lookup)
+
+    monkeypatch.setattr(proxy_utils, "create_model_info_response", _resolved)
+
+    with auth_as():
+        response = client.get(path, headers={"anthropic-version": "2023-06-01"})
+
+    assert response.status_code == 200
+    gpt_4, claude = response.json()["data"]
+    assert (gpt_4["max_input_tokens"], gpt_4["max_tokens"]) == (300000, 32000)
+    assert (claude["max_input_tokens"], claude["max_tokens"]) == (500000, 4096)
 
 
 @pytest.mark.parametrize("path", ["/v1/models", "/models"])