fix(vertex_ai): keep context-1m beta out of the Vertex anthropic-beta header

Vertex's rawPredict validates the anthropic-beta HTTP header against an allowed
set and returns a 400 ("Unexpected value(s) context-1m-2025-08-07 for the
anthropic-beta header") when context-1m-2025-08-07 is sent there. Anthropic betas
on Vertex are read from the anthropic_beta request body, so that beta has to travel
in the body only.

VertexAIAnthropicConfig.transform_request was forwarding every beta into both the
body and the header (#22321). That is correct for betas Vertex wants in the header
(tool-search, context_management) but it breaks 1M context: any request asking for
context-1m on a Vertex Claude deployment fails with a 400, and a fallback that lands
on Vertex carrying the [1m] beta dies the same way.

This keeps body-only betas (currently context-1m-2025-08-07) in the anthropic_beta
body field while excluding them from the anthropic-beta header

litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/transformation.py

@@ -12,6 +12,10 @@ from litellm.types.utils import ModelResponse
 from ....anthropic.chat.transformation import AnthropicConfig
 from .output_params_utils import sanitize_vertex_anthropic_output_params
 
+# Betas that Vertex's rawPredict rejects (400) when sent in the anthropic-beta HTTP
+# header; they are only accepted in the anthropic_beta request body (e.g. context-1m).
+VERTEX_BODY_ONLY_ANTHROPIC_BETAS = frozenset({"context-1m-2025-08-07"})
+
 
 class VertexAIError(Exception):
     def __init__(self, status_code, message):
@@ -137,7 +141,9 @@ class VertexAIAnthropicConfig(AnthropicConfig):
 
         if beta_set:
             data["anthropic_beta"] = list(beta_set)
-            headers["anthropic-beta"] = ",".join(beta_set)
+            header_betas = beta_set - VERTEX_BODY_ONLY_ANTHROPIC_BETAS
+            if header_betas:
+                headers["anthropic-beta"] = ",".join(header_betas)
 
         return data
 

tests/test_litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_partner_models_anthropic_transformation.py

@@ -463,6 +463,62 @@ def test_vertex_ai_anthropic_no_extra_headers_unchanged():
     assert "anthropic-beta" not in headers
 
 
+def test_vertex_ai_anthropic_context_1m_beta_in_body_not_header():
+    """context-1m-2025-08-07 must go in the anthropic_beta body field only.
+
+    Vertex's rawPredict 400s ("Unexpected value(s) ... for the anthropic-beta header")
+    when this beta is sent in the HTTP header, so it must not be forwarded there.
+    """
+    config = VertexAIAnthropicConfig()
+
+    headers = {}
+    optional_params = {
+        "max_tokens": 100,
+        "is_vertex_request": True,
+        "extra_headers": {"anthropic-beta": "context-1m-2025-08-07"},
+    }
+
+    result = config.transform_request(
+        model="claude-sonnet-4-5@20250929",
+        messages=[{"role": "user", "content": "Hello"}],
+        optional_params=optional_params,
+        litellm_params={},
+        headers=headers,
+    )
+
+    assert "context-1m-2025-08-07" in result["anthropic_beta"]
+    assert "anthropic-beta" not in headers
+
+
+def test_vertex_ai_anthropic_context_1m_excluded_from_header_but_others_kept():
+    """When context-1m is combined with a header-compatible beta, only context-1m is
+    stripped from the HTTP header; both still travel in the body."""
+    config = VertexAIAnthropicConfig()
+
+    headers = {}
+    optional_params = {
+        "max_tokens": 100,
+        "is_vertex_request": True,
+        "extra_headers": {
+            "anthropic-beta": "context-1m-2025-08-07,interleaved-thinking-2025-05-14",
+        },
+    }
+
+    result = config.transform_request(
+        model="claude-sonnet-4-5@20250929",
+        messages=[{"role": "user", "content": "Hello"}],
+        optional_params=optional_params,
+        litellm_params={},
+        headers=headers,
+    )
+
+    assert "context-1m-2025-08-07" in result["anthropic_beta"]
+    assert "interleaved-thinking-2025-05-14" in result["anthropic_beta"]
+
+    assert "interleaved-thinking-2025-05-14" in headers["anthropic-beta"]
+    assert "context-1m-2025-08-07" not in headers["anthropic-beta"]
+
+
 def test_vertex_ai_partner_models_anthropic_remove_prompt_caching_scope_beta_header():
     """
     Test that remove_unsupported_beta correctly filters out prompt-caching-scope-2026-01-05