fix(azure_ai): forward the thinking-binding-controls beta header to Foundry (#48033)

litellm-rust/crates/llms-types/src/providers/anthropic/beta.rs

@@ -235,7 +235,10 @@ impl AnthropicBeta {
                 BetaProvider::Databricks,
             ],
             Self::Compact20260112 => &[BetaProvider::AzureAi, BetaProvider::BedrockConverse],
-            Self::ComputerUse20250124 | Self::ComputerUse20251124 | Self::Context1m20250807 => &[],
+            Self::ComputerUse20250124
+            | Self::ComputerUse20251124
+            | Self::Context1m20250807
+            | Self::ThinkingBindingControls20260801 => &[],
             Self::DangerousToolUse20260903 => &[BetaProvider::BedrockConverse],
             Self::Effort20251124 => &[BetaProvider::VertexAi],
             Self::FastMode20260201 | Self::Oauth20250420 => &[
@@ -272,7 +275,6 @@ impl AnthropicBeta {
                 BetaProvider::Databricks,
             ],
             Self::StructuredOutputs20251113 => &[BetaProvider::Bedrock, BetaProvider::VertexAi],
-            Self::ThinkingBindingControls20260801 => &[BetaProvider::AzureAi],
             Self::ThinkingDisplayUpdates20260818 => &[BetaProvider::Databricks],
             Self::ToolSearchTool20251019 => &[
                 BetaProvider::Anthropic,

litellm/anthropic_beta_headers_config.json

@@ -67,7 +67,7 @@
     "structured-outputs-2025-11-13": "structured-outputs-2025-11-13",
     "text_editor_20241022": null,
     "text_editor_20250124": null,
-    "thinking-binding-controls-2026-08-01": null,
+    "thinking-binding-controls-2026-08-01": "thinking-binding-controls-2026-08-01",
     "thinking-display-updates-2026-08-18": "thinking-display-updates-2026-08-18",
     "token-efficient-tools-2025-02-19": null,
     "web-fetch-2025-09-10": "web-fetch-2025-09-10",

litellm/proxy/litellm_pre_call_utils.py

@@ -4278,6 +4278,9 @@ _ANTHROPIC_API_HEADER_PROVIDERS: Final = ",".join(
         LlmProviders.VERTEX_AI.value,
     )
 )
+_NATIVE_MESSAGES_ANTHROPIC_API_HEADER_PROVIDERS: Final = ",".join(
+    (LlmProviders.AZURE_AI.value, LlmProviders.DATABRICKS.value)
+)
 _ANTHROPIC_OAUTH_CREDENTIAL_PROVIDERS: Final = ",".join(sorted(ANTHROPIC_OAUTH_FORWARD_PROVIDERS))
 
 
@@ -4306,7 +4309,7 @@ def add_provider_specific_headers_to_request(
                 custom_llm_provider=_ANTHROPIC_API_HEADER_PROVIDERS, extra_headers=anthropic_api_headers
             ),
             ProviderSpecificHeader(
-                custom_llm_provider=LlmProviders.DATABRICKS.value,
+                custom_llm_provider=_NATIVE_MESSAGES_ANTHROPIC_API_HEADER_PROVIDERS,
                 extra_headers=anthropic_api_headers,
                 native_anthropic_messages_only=True,
             ),

tests/unit/proxy/test_litellm_pre_call_utils.py

@@ -8111,6 +8111,67 @@ def test_client_anthropic_api_headers_stay_off_openai_compatible_providers():
     assert forwarded == {}
 
 
+def test_client_anthropic_api_headers_stay_off_azure_ai_outside_native_messages():
+    forwarded = _headers_forwarded_to({"anthropic-beta": "thinking-binding-controls-2026-08-01"}, "azure_ai")
+
+    assert forwarded == {}
+
+
+@pytest.mark.asyncio
+@pytest.mark.usefixtures("local_beta_headers_config")
+async def test_client_thinking_binding_beta_reaches_azure_ai_foundry_claude():
+    import httpx
+
+    from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
+
+    data: Final[dict] = {"model": "claude-opus-4-8"}  # mutable-ok: the function under test writes into it
+    add_provider_specific_headers_to_request(
+        data=data,
+        headers={
+            "content-type": "application/json",
+            "anthropic-beta": "thinking-binding-controls-2026-08-01",
+            "anthropic-version": "2026-08-01",
+        },
+    )
+    received: Final[list[httpx.Request]] = []  # mutable-ok: the fake Foundry endpoint records what the proxy sent
+
+    def foundry_answers(request: httpx.Request) -> httpx.Response:
+        received.append(request)
+        return httpx.Response(
+            status_code=200,
+            json={
+                "id": "msg_01",
+                "type": "message",
+                "role": "assistant",
+                "model": "claude-opus-4-8",
+                "content": [{"type": "text", "text": "hello"}],
+                "stop_reason": "end_turn",
+                "stop_sequence": None,
+                "usage": {"input_tokens": 12, "output_tokens": 6},
+            },
+            request=request,
+        )
+
+    upstream: Final = AsyncHTTPHandler(transport=httpx.MockTransport(foundry_answers))
+    try:
+        await litellm.anthropic.messages.acreate(
+            max_tokens=16,
+            messages=[{"role": "user", "content": "hi"}],
+            thinking={"type": "adaptive", "block_binding": {"prefix_mismatch_behavior": "drop_block"}},
+            model="azure_ai/claude-opus-4-8",
+            api_key="foundry-key",
+            api_base="https://my-resource.services.ai.azure.com/anthropic",
+            client=upstream,
+            provider_specific_header=data["provider_specific_header"],
+        )
+    finally:
+        await upstream.close()
+
+    assert received[0].headers["anthropic-beta"] == "thinking-binding-controls-2026-08-01"
+    assert received[0].headers["anthropic-version"] == "2026-08-01"
+    assert json.loads(received[0].content)["thinking"]["block_binding"] == {"prefix_mismatch_behavior": "drop_block"}
+
+
 TYPESAFE_DANGER_ZONE_HEADERS = {
     "TypeSafe-Danger-Zone-Images": "true",
     "typesafe-danger-zone-long-context": "true",

tests/unit/test_anthropic_beta_headers_filtering.py

@@ -429,13 +429,13 @@ class TestAnthropicBetaHeadersFiltering:
         assert filtered == ["fine-grained-tool-streaming-2025-05-14"]
 
     @pytest.mark.parametrize(
-        "provider", ["anthropic", "bedrock", "bedrock_converse", "vertex_ai", "databricks"]
+        "provider", ["anthropic", "azure_ai", "bedrock", "bedrock_converse", "vertex_ai", "databricks"]
     )
     def test_thinking_binding_controls_forwarded(self, provider):
         """`thinking.block_binding` (preserved thinking, Claude Fable 5.1) is only
         accepted alongside thinking-binding-controls-2026-08-01. The body field is
         forwarded untouched, so stripping the header (previously unknown, hence
-        dropped) makes Bedrock and Vertex reject the request with
+        dropped) makes Bedrock, Vertex and Azure AI Foundry reject the request with
         "thinking.adaptive.block_binding: Extra inputs are not permitted"."""
         filtered = filter_and_transform_beta_headers(
             beta_headers=["thinking-binding-controls-2026-08-01"],