fix(bedrock): clamp thinking.budget_tokens to minimum 1024

Bedrock rejects thinking.budget_tokens values below 1024 with a 400
error. This adds automatic clamping in the LiteLLM transformation
layer so callers (e.g. router with reasoning_effort="low") don't
need to know about the provider-specific minimum.

Fixes #21297

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

litellm/constants.py

@@ -319,6 +319,9 @@ NON_LLM_CONNECTION_TIMEOUT = int(
 MAX_EXCEPTION_MESSAGE_LENGTH = int(os.getenv("MAX_EXCEPTION_MESSAGE_LENGTH", 2000))
 MAX_STRING_LENGTH_PROMPT_IN_DB = int(os.getenv("MAX_STRING_LENGTH_PROMPT_IN_DB", 2048))
 BEDROCK_MAX_POLICY_SIZE = int(os.getenv("BEDROCK_MAX_POLICY_SIZE", 75))
+BEDROCK_MIN_THINKING_BUDGET_TOKENS = int(
+    os.getenv("BEDROCK_MIN_THINKING_BUDGET_TOKENS", 1024)
+)
 REPLICATE_POLLING_DELAY_SECONDS = float(
     os.getenv("REPLICATE_POLLING_DELAY_SECONDS", 0.5)
 )

litellm/llms/bedrock/chat/converse_transformation.py

@@ -11,7 +11,10 @@ import httpx
 
 import litellm
 from litellm._logging import verbose_logger
-from litellm.constants import RESPONSE_FORMAT_TOOL_NAME
+from litellm.constants import (
+    BEDROCK_MIN_THINKING_BUDGET_TOKENS,
+    RESPONSE_FORMAT_TOOL_NAME,
+)
 from litellm.litellm_core_utils.core_helpers import (
     filter_exceptions_from_params,
     filter_internal_params,
@@ -434,6 +437,25 @@ class AmazonConverseConfig(BaseConfig):
                 reasoning_effort=reasoning_effort, model=model
             )
 
+    @staticmethod
+    def _clamp_thinking_budget_tokens(optional_params: dict) -> None:
+        """
+        Clamp thinking.budget_tokens to the Bedrock minimum (1024).
+
+        Bedrock returns a 400 error if budget_tokens < 1024.
+        """
+        thinking = optional_params.get("thinking")
+        if isinstance(thinking, dict):
+            budget = thinking.get("budget_tokens")
+            if isinstance(budget, int) and budget < BEDROCK_MIN_THINKING_BUDGET_TOKENS:
+                verbose_logger.debug(
+                    "Bedrock requires thinking.budget_tokens >= %d, got %d. "
+                    "Clamping to minimum.",
+                    BEDROCK_MIN_THINKING_BUDGET_TOKENS,
+                    budget,
+                )
+                thinking["budget_tokens"] = BEDROCK_MIN_THINKING_BUDGET_TOKENS
+
     def get_supported_openai_params(self, model: str) -> List[str]:
         from litellm.utils import supports_function_calling
 
@@ -871,9 +893,14 @@ class AmazonConverseConfig(BaseConfig):
         Checks 'non_default_params' for 'thinking' and 'max_tokens'
 
         if 'thinking' is enabled and 'max_tokens' is not specified, set 'max_tokens' to the thinking token budget + DEFAULT_MAX_TOKENS
+
+        Also clamps thinking.budget_tokens to the Bedrock minimum (1024) to
+        prevent 400 errors from the Bedrock API.
         """
         from litellm.constants import DEFAULT_MAX_TOKENS
 
+        self._clamp_thinking_budget_tokens(optional_params)
+
         is_thinking_enabled = self.is_thinking_enabled(optional_params)
         is_max_tokens_in_request = self.is_max_tokens_in_request(non_default_params)
         if is_thinking_enabled and not is_max_tokens_in_request:

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

@@ -2934,3 +2934,47 @@ def test_drop_thinking_param_when_thinking_blocks_missing():
     finally:
         # Restore original modify_params setting
         litellm.modify_params = original_modify_params
+
+
+class TestBedrockMinThinkingBudgetTokens:
+    """Test that thinking.budget_tokens is clamped to the Bedrock minimum (1024)."""
+
+    def _map_params(
+        self, thinking_value, model="anthropic.claude-3-7-sonnet-20250219-v1:0"
+    ):
+        """Helper to call map_openai_params with the given thinking value."""
+        config = AmazonConverseConfig()
+        non_default_params = {"thinking": thinking_value}
+        optional_params = {"thinking": thinking_value}
+        return config.map_openai_params(
+            non_default_params=non_default_params,
+            optional_params=optional_params,
+            model=model,
+            drop_params=False,
+        )
+
+    def test_budget_tokens_below_minimum_is_clamped(self):
+        """budget_tokens < 1024 should be clamped to 1024."""
+        result = self._map_params({"type": "enabled", "budget_tokens": 499})
+        assert result["thinking"]["budget_tokens"] == 1024
+
+    def test_budget_tokens_at_minimum_is_unchanged(self):
+        """budget_tokens == 1024 should remain 1024."""
+        result = self._map_params({"type": "enabled", "budget_tokens": 1024})
+        assert result["thinking"]["budget_tokens"] == 1024
+
+    def test_budget_tokens_above_minimum_is_unchanged(self):
+        """budget_tokens > 1024 should remain unchanged."""
+        result = self._map_params({"type": "enabled", "budget_tokens": 2048})
+        assert result["thinking"]["budget_tokens"] == 2048
+
+    def test_no_thinking_param_does_not_error(self):
+        """When thinking is not provided, map_openai_params should not raise."""
+        config = AmazonConverseConfig()
+        result = config.map_openai_params(
+            non_default_params={},
+            optional_params={},
+            model="anthropic.claude-3-7-sonnet-20250219-v1:0",
+            drop_params=False,
+        )
+        assert "thinking" not in result or result.get("thinking") is None