fix(bedrock): clamp reasoning_effort thinking budget to max_tokens on converse

Fixes #39627. _handle_reasoning_effort_parameter assigned the mapped
thinking budget without capping it against max_tokens, so any request
with max_tokens <= the mapped budget (e.g. medium's 2048) got a 400
from Bedrock. Now routes through AnthropicConfig.cap_thinking_budget_to_max_tokens,
matching the existing adaptive-downgrade branch and the /v1/messages fix in 71a9516.

litellm/llms/bedrock/chat/converse_transformation.py

@@ -418,7 +418,9 @@ class AmazonConverseConfig(BaseConfig):
             }
         }
 
-    def _handle_reasoning_effort_parameter(self, model: str, reasoning_effort: str, optional_params: dict) -> None:
+    def _handle_reasoning_effort_parameter(
+        self, model: str, reasoning_effort: str, optional_params: dict, max_tokens: int | None = None
+    ) -> None:
         """
         Handle the reasoning_effort parameter based on the model type.
 
@@ -443,6 +445,8 @@ class AmazonConverseConfig(BaseConfig):
                 custom_llm_provider="bedrock",
                 llm_provider="bedrock_converse",
             )
+            if mapped_thinking is not None:
+                mapped_thinking = AnthropicConfig.cap_thinking_budget_to_max_tokens(mapped_thinking, max_tokens)
             if mapped_thinking is None:
                 optional_params.pop("thinking", None)
                 optional_params.pop("output_config", None)
@@ -948,7 +952,10 @@ class AmazonConverseConfig(BaseConfig):
                     )
             elif param == "reasoning_effort" and isinstance(value, str):
                 self._handle_reasoning_effort_parameter(
-                    model=model, reasoning_effort=value, optional_params=optional_params
+                    model=model,
+                    reasoning_effort=value,
+                    optional_params=optional_params,
+                    max_tokens=non_default_params.get("max_completion_tokens") or non_default_params.get("max_tokens"),
                 )
             elif param == "output_config" and isinstance(value, dict):
                 mapped_output_config = dict(value)

tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py

@@ -462,6 +462,40 @@ def test_reasoning_effort_none_omits_thinking_for_anthropic_converse(model):
     assert "thinking" not in optional_params
 
 
+@pytest.mark.parametrize(
+    "max_tokens,expect_thinking,expect_budget_below_max_tokens",
+    [
+        (1024, False, None),  # at/below Bedrock's min thinking budget -> thinking dropped entirely
+        (2048, True, True),   # medium's default budget (2048) equals max_tokens -> must be clamped below it
+        (4096, True, False),  # budget already fits comfortably under max_tokens -> left unchanged
+    ],
+)
+def test_reasoning_effort_thinking_budget_clamped_to_max_tokens_converse(
+    max_tokens, expect_thinking, expect_budget_below_max_tokens
+):
+    """Regression #39627: a deployment-level reasoning_effort must not send a
+    thinking.budget_tokens that is >= max_tokens on Bedrock Converse. Bedrock
+    requires maxTokens strictly greater than thinking.budget_tokens, so the
+    mapped budget must be clamped the same way the adaptive-downgrade branch
+    already clamps it, or dropped when even the 1024 minimum can't fit."""
+    config = AmazonConverseConfig()
+
+    optional_params = config.map_openai_params(
+        non_default_params={"max_tokens": max_tokens, "reasoning_effort": "medium"},
+        optional_params={},
+        model="us.anthropic.claude-haiku-4-5-20251001-v1:0",
+        drop_params=False,
+    )
+
+    if not expect_thinking:
+        assert optional_params.get("thinking") is None
+        return
+
+    thinking = optional_params["thinking"]
+    budget = thinking["budget_tokens"]
+    assert budget < max_tokens if expect_budget_below_max_tokens else budget == 2048
+
+
 @pytest.mark.parametrize(
     "model,effort,expected_effort",
     [