Merge pull request #39159 from BerriAI/litellm_bedrock_converse_legacy_thinking_adaptive

fix(anthropic): upgrade legacy thinking to adaptive on adaptive-only Claude models for chat, Bedrock Converse, Invoke, Vertex AI, and Databricks

litellm/llms/anthropic/chat/transformation.py

@@ -1565,6 +1565,9 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
                         optional_params.pop("thinking", None)
                 else:
                     optional_params["thinking"] = value
+                    AnthropicModelInfo.translate_legacy_thinking_for_adaptive_model(
+                        model=model, optional_params=optional_params, custom_llm_provider=self._resolved_provider
+                    )
             elif param == "reasoning_effort":
                 # Accept both string ("low") and dict ({"effort": "low",
                 # "summary": "concise"}). The Responses->Chat parser keeps the

litellm/llms/anthropic/common_utils.py

@@ -13,7 +13,12 @@ import httpx
 from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError
 
 import litellm
-from litellm.constants import DEFAULT_MODEL_CREATED_AT_TIME
+from litellm.constants import (
+    DEFAULT_MODEL_CREATED_AT_TIME,
+    DEFAULT_REASONING_EFFORT_HIGH_THINKING_BUDGET,
+    DEFAULT_REASONING_EFFORT_MEDIUM_THINKING_BUDGET,
+    DEFAULT_REASONING_EFFORT_XHIGH_THINKING_BUDGET,
+)
 from litellm.litellm_core_utils.prompt_templates.common_utils import (
     get_file_ids_from_messages,
 )
@@ -534,6 +539,51 @@ class AnthropicModelInfo(BaseLLMModelInfo):
         )
         optional_params.pop("thinking", None)
 
+    @staticmethod
+    def translate_legacy_thinking_for_adaptive_model(
+        model: str,
+        optional_params: MutableMapping[str, object],  # mutable-ok: in-place out-param like the sibling helpers
+        custom_llm_provider: str,
+    ) -> None:
+        """Translate legacy ``thinking.type=enabled`` to adaptive for the
+        adaptive-thinking models that reject it (4.7+ and the 5 families).
+        Models flagged ``supports_legacy_thinking`` (the 4.6 family) accept the
+        legacy shape natively, so it is forwarded verbatim and the caller's
+        ``budget_tokens`` cap keeps applying. Caller-provided
+        ``output_config.effort`` is never overridden.
+        """
+        if not AnthropicModelInfo._is_adaptive_thinking_model(model, custom_llm_provider):
+            return
+        if AnthropicModelInfo._supports_legacy_thinking(model, custom_llm_provider):
+            return
+        thinking: Final = optional_params.get("thinking")
+        if not isinstance(thinking, dict) or thinking.get("type") != "enabled":
+            return
+
+        effort: Final = AnthropicModelInfo._legacy_budget_to_effort(
+            model=model,
+            budget_tokens=int(thinking.get("budget_tokens") or 0),
+            custom_llm_provider=custom_llm_provider,
+        )
+        existing_output_config: Final = optional_params.get("output_config")
+        optional_params["thinking"] = {"type": "adaptive"}
+        optional_params["output_config"] = {
+            "effort": effort,
+            **(existing_output_config if isinstance(existing_output_config, dict) else MappingProxyType({})),
+        }
+
+    @staticmethod
+    def _legacy_budget_to_effort(model: str, budget_tokens: int, custom_llm_provider: str) -> str:
+        if budget_tokens >= DEFAULT_REASONING_EFFORT_XHIGH_THINKING_BUDGET and (
+            AnthropicModelInfo._supports_model_capability(model, "supports_xhigh_reasoning_effort", custom_llm_provider)
+        ):
+            return "xhigh"
+        if budget_tokens >= DEFAULT_REASONING_EFFORT_HIGH_THINKING_BUDGET:
+            return "high"
+        if budget_tokens >= DEFAULT_REASONING_EFFORT_MEDIUM_THINKING_BUDGET:
+            return "medium"
+        return "low"
+
     def is_effort_used(
         self,
         optional_params: dict | None,

litellm/llms/anthropic/experimental_pass_through/messages/transformation.py

@@ -3,11 +3,6 @@ from typing import Any, Final
 
 import httpx
 
-from litellm.constants import (
-    DEFAULT_REASONING_EFFORT_HIGH_THINKING_BUDGET,
-    DEFAULT_REASONING_EFFORT_MEDIUM_THINKING_BUDGET,
-    DEFAULT_REASONING_EFFORT_XHIGH_THINKING_BUDGET,
-)
 from litellm.exceptions import AuthenticationError
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
 from litellm.litellm_core_utils.litellm_logging import verbose_logger
@@ -400,46 +395,6 @@ class AnthropicMessagesConfig(BaseAnthropicMessagesConfig):
             existing_output_config.setdefault("effort", mapped_effort)
             optional_params["output_config"] = existing_output_config
 
-    @staticmethod
-    def _translate_legacy_thinking_for_adaptive_model(
-        model: str, optional_params: dict, custom_llm_provider: str
-    ) -> None:
-        """Translate legacy ``thinking.type=enabled`` to adaptive for the
-        adaptive-thinking models that reject it (4.7+ and the 5 families).
-        Models flagged ``supports_legacy_thinking`` (the 4.6 family) accept the
-        legacy shape natively, so it is forwarded verbatim and the caller's
-        ``budget_tokens`` cap keeps applying. Caller-provided
-        ``output_config.effort`` is never overridden.
-        """
-        from litellm.llms.anthropic.chat.transformation import AnthropicConfig
-
-        if not AnthropicModelInfo._is_adaptive_thinking_model(model, custom_llm_provider):
-            return
-        if AnthropicModelInfo._supports_legacy_thinking(model, custom_llm_provider):
-            return
-        thinking: Final = optional_params.get("thinking")
-        if not isinstance(thinking, dict) or thinking.get("type") != "enabled":
-            return
-
-        budget: Final = int(thinking.get("budget_tokens") or 0)
-        if budget >= DEFAULT_REASONING_EFFORT_XHIGH_THINKING_BUDGET and (
-            AnthropicConfig._supports_effort_level(model, "xhigh", custom_llm_provider)
-        ):
-            effort = "xhigh"
-        elif budget >= DEFAULT_REASONING_EFFORT_HIGH_THINKING_BUDGET:
-            effort = "high"
-        elif budget >= DEFAULT_REASONING_EFFORT_MEDIUM_THINKING_BUDGET:
-            effort = "medium"
-        else:
-            effort = "low"
-
-        optional_params["thinking"] = {"type": "adaptive"}
-        existing_output_config = optional_params.get("output_config")
-        if not isinstance(existing_output_config, dict):
-            existing_output_config = {}
-        existing_output_config.setdefault("effort", effort)
-        optional_params["output_config"] = existing_output_config
-
     @staticmethod
     def _translate_adaptive_effort_for_non_adaptive_model(
         model: str, optional_params: dict, max_tokens: int | None, custom_llm_provider: str
@@ -606,7 +561,7 @@ class AnthropicMessagesConfig(BaseAnthropicMessagesConfig):
             custom_llm_provider=self._resolved_provider,
         )
 
-        self._translate_legacy_thinking_for_adaptive_model(
+        AnthropicModelInfo.translate_legacy_thinking_for_adaptive_model(
             model=model,
             optional_params=anthropic_messages_optional_request_params,
             custom_llm_provider=self._resolved_provider,

litellm/llms/azure_ai/anthropic/transformation.py

@@ -17,13 +17,14 @@ def _promote_extra_body_to_optional_params(optional_params: dict) -> None:
     ``output_config`` get auto-routed into ``extra_body`` by
     ``add_provider_specific_params_to_optional_params``. For the Azure→Anthropic
     route those keys must reach the request body and be validated, so promote
-    them. ``setdefault`` keeps explicit top-level values authoritative.
+    them. The caller's values overwrite mapped top-level duplicates, matching
+    the native ``anthropic`` provider, where the same passthrough lands on
+    top-level ``optional_params`` after mapping.
     """
     extra_body: Final = optional_params.get("extra_body")
     if not isinstance(extra_body, dict) or not extra_body:
         return
-    for k, v in extra_body.items():
-        optional_params.setdefault(k, v)
+    optional_params.update(extra_body)
     optional_params.pop("extra_body", None)
 
 

litellm/llms/bedrock/chat/converse_transformation.py

@@ -943,6 +943,9 @@ class AmazonConverseConfig(BaseConfig):
                         litellm.verbose_logger.warning(DROP_UNSUPPORTED_ADAPTIVE_THINKING_WARNING, model)
                 else:
                     optional_params["thinking"] = value
+                    AnthropicModelInfo.translate_legacy_thinking_for_adaptive_model(
+                        model=model, optional_params=optional_params, custom_llm_provider="bedrock"
+                    )
             elif param == "reasoning_effort" and isinstance(value, str):
                 self._handle_reasoning_effort_parameter(
                     model=model, reasoning_effort=value, optional_params=optional_params

litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py

@@ -107,6 +107,10 @@ class AmazonAnthropicClaudeConfig(AmazonInvokeConfig, AnthropicConfig):
         # Restore original model name
         model = original_model
 
+        AnthropicModelInfo.translate_legacy_thinking_for_adaptive_model(
+            model=original_model, optional_params=optional_params, custom_llm_provider="bedrock"
+        )
+
         # The stub model hides the original model from the parent's forced-tool-use backstop
         response_format_tool_choice: Final = optional_params.get("tool_choice")
         if (

litellm/llms/databricks/chat/transformation.py

@@ -330,6 +330,10 @@ class DatabricksConfig(DatabricksBase, OpenAILikeChatConfig, AnthropicConfig):
     ) -> dict:
         is_thinking_enabled: Final = self.is_thinking_enabled(non_default_params)
         mapped_params: Final = super().map_openai_params(non_default_params, optional_params, model, drop_params)
+        if "claude" in model:
+            AnthropicConfig.translate_legacy_thinking_for_adaptive_model(
+                model=model, optional_params=mapped_params, custom_llm_provider="databricks"
+            )
         if "tools" in mapped_params:
             mapped_params["tools"] = self._map_openai_to_dbrx_tool(model=model, tools=mapped_params["tools"])
         if "max_completion_tokens" in non_default_params and replace_max_completion_tokens_with_max_tokens:

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

@@ -177,6 +177,10 @@ class VertexAIAnthropicConfig(AnthropicConfig):
         # Restore original model name for any other processing
         model = original_model
 
+        AnthropicModelInfo.translate_legacy_thinking_for_adaptive_model(
+            model=original_model, optional_params=optional_params, custom_llm_provider="vertex_ai"
+        )
+
         return optional_params
 
     def transform_response(

tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py

@@ -3127,6 +3127,31 @@ def test_reasoning_effort_accepts_dict_shape_for_non_adaptive_model(
     )
 
 
+@pytest.mark.parametrize(
+    "model,budget_tokens,expected",
+    [
+        ("claude-opus-4-8", 4096, ({"type": "adaptive"}, {"effort": "high"})),
+        ("claude-opus-4-7", 24000, ({"type": "adaptive"}, {"effort": "xhigh"})),
+        ("claude-opus-4-6", 4096, ({"type": "enabled", "budget_tokens": 4096}, None)),
+        ("claude-sonnet-4-5-20250929", 4096, ({"type": "enabled", "budget_tokens": 4096}, None)),
+    ],
+)
+def test_legacy_thinking_translated_to_adaptive_on_adaptive_only_models(model, budget_tokens, expected):
+    """Adaptive-only models reject thinking={type: enabled} with a 400, so the
+    legacy shape must be upgraded to adaptive + output_config.effort on
+    /chat/completions too, while models that accept it keep the caller's budget."""
+    config = AnthropicConfig()
+
+    result = config.map_openai_params(
+        non_default_params={"thinking": {"type": "enabled", "budget_tokens": budget_tokens}, "max_tokens": 64000},
+        optional_params={},
+        model=model,
+        drop_params=False,
+    )
+
+    assert (result["thinking"], result.get("output_config")) == expected
+
+
 @pytest.mark.parametrize(
     "bad_value",
     [

tests/test_litellm/llms/azure_ai/claude/test_azure_anthropic_transformation.py

@@ -362,8 +362,8 @@ class TestAzureAnthropicConfig:
             )
         assert "xhigh" in str(exc_info.value)
 
-    def test_extra_body_promotion_does_not_clobber_top_level(self):
-        """Top-level ``optional_params`` wins over duplicates in ``extra_body``."""
+    def test_extra_body_promotion_overrides_mapped_top_level(self):
+        """The caller's ``extra_body`` wins over a mapped top-level duplicate, like the native ``anthropic`` passthrough."""
         config = AzureAnthropicConfig()
 
         messages = [{"role": "user", "content": "Hello"}]
@@ -383,7 +383,31 @@ class TestAzureAnthropicConfig:
             headers=headers,
         )
 
-        assert result["output_config"] == {"effort": "low"}
+        assert result["output_config"] == {"effort": "high"}
+
+    def test_legacy_thinking_upgrade_keeps_caller_effort_from_extra_body(self, local_model_cost_map):
+        config = AzureAnthropicConfig()
+
+        mapped = config.map_openai_params(
+            non_default_params={"thinking": {"type": "enabled", "budget_tokens": 1024}, "max_tokens": 100},
+            optional_params={},
+            model="claude-opus-4-8",
+            drop_params=False,
+        )
+        assert mapped["thinking"] == {"type": "adaptive"}
+        assert mapped["output_config"] == {"effort": "low"}
+
+        result = config.transform_request(
+            model="claude-opus-4-8",
+            messages=[{"role": "user", "content": "Hello"}],
+            optional_params={**mapped, "extra_body": {"output_config": {"effort": "high"}}},
+            litellm_params={"api_key": "test-key"},
+            headers={"api-key": "test-key", "anthropic-version": "2023-06-01"},
+        )
+
+        assert result["thinking"] == {"type": "adaptive"}
+        assert result["output_config"] == {"effort": "high"}
+        assert "extra_body" not in result
 
     def test_context_management_mixed_edits_beta_headers(self):
         """Test that context_management with both compact and other edits adds both beta headers"""

tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py

@@ -671,3 +671,46 @@ def test_bedrock_chat_invoke_fable_5_1_response_format_avoids_forced_tool_choice
     assert "output_format" not in result
     assert "tools" in result
     assert "tool_choice" not in result
+
+
+@pytest.mark.parametrize("model", ["us.anthropic.claude-sonnet-5", "us.anthropic.claude-fable-5-1"])
+def test_bedrock_chat_invoke_tool_based_response_format_still_upgrades_legacy_thinking(local_model_cost_map, model):
+    result = AmazonAnthropicClaudeConfig().map_openai_params(
+        non_default_params={
+            "response_format": {
+                "type": "json_schema",
+                "json_schema": {
+                    "name": "test_schema",
+                    "schema": {"type": "object", "properties": {"result": {"type": "string"}}},
+                },
+            },
+            "thinking": {"type": "enabled", "budget_tokens": 4096},
+            "max_tokens": 8192,
+        },
+        optional_params={},
+        model=model,
+        drop_params=False,
+    )
+
+    assert "tools" in result
+    assert result["thinking"] == {"type": "adaptive"}
+    assert result["output_config"] == {"effort": "high"}
+
+
+def test_bedrock_chat_invoke_response_format_stub_still_upgrades_legacy_thinking(local_model_cost_map):
+    """Regression: the tool-based ``response_format`` path swaps in a Claude 3 stub
+    model before the shared Anthropic mapping, which hid the adaptive-only model
+    from the legacy ``thinking`` upgrade and left ``type=enabled`` on the wire."""
+    result = AmazonAnthropicClaudeConfig().map_openai_params(
+        non_default_params={
+            "response_format": {"type": "json_object"},
+            "thinking": {"type": "enabled", "budget_tokens": 4096},
+            "max_tokens": 8192,
+        },
+        optional_params={},
+        model="us.anthropic.claude-fable-5-1",
+        drop_params=False,
+    )
+
+    assert result["thinking"] == {"type": "adaptive"}
+    assert result["output_config"] == {"effort": "high"}

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

@@ -6347,6 +6347,82 @@ def test_adaptive_thinking_passes_through_on_46_plus_converse(model):
     assert optional_params.get("thinking") == {"type": "adaptive"}
 
 
+@pytest.mark.parametrize(
+    "model,budget_tokens,expected_effort",
+    [
+        ("anthropic.claude-opus-4-8", 4096, "high"),
+        ("us.anthropic.claude-opus-4-8", 2000, "low"),
+        ("global.anthropic.claude-opus-4-8", 12000, "xhigh"),
+        ("us.anthropic.claude-opus-4-7", 3000, "medium"),
+        ("anthropic.claude-fable-5", 4096, "high"),
+    ],
+)
+def test_legacy_thinking_translated_to_adaptive_on_adaptive_only_converse(model, budget_tokens, expected_effort):
+    """Adaptive-only models (4.7+, 5 families) reject thinking={type: enabled}
+    with a 400 on Bedrock Converse, so the legacy shape from callers like Claude
+    Code must be upgraded to thinking={type: adaptive} + output_config.effort
+    derived from budget_tokens, matching the /v1/messages passthrough."""
+    config = AmazonConverseConfig()
+
+    optional_params = config.map_openai_params(
+        non_default_params={"thinking": {"type": "enabled", "budget_tokens": budget_tokens}, "max_tokens": 64000},
+        optional_params={},
+        model=model,
+        drop_params=False,
+    )
+    request = config.transform_request(
+        model=model,
+        messages=[{"role": "user", "content": "hi"}],
+        optional_params=optional_params,
+        litellm_params={},
+        headers={},
+    )
+
+    assert request["additionalModelRequestFields"]["thinking"] == {"type": "adaptive"}
+    assert request["additionalModelRequestFields"]["output_config"] == {"effort": expected_effort}
+
+
+def test_legacy_thinking_translation_keeps_caller_output_config_effort_converse():
+    config = AmazonConverseConfig()
+
+    optional_params = config.map_openai_params(
+        non_default_params={
+            "output_config": {"effort": "low"},
+            "thinking": {"type": "enabled", "budget_tokens": 12000},
+            "max_tokens": 64000,
+        },
+        optional_params={},
+        model="anthropic.claude-opus-4-8",
+        drop_params=False,
+    )
+
+    assert optional_params["thinking"] == {"type": "adaptive"}
+    assert optional_params["output_config"] == {"effort": "low"}
+
+
+@pytest.mark.parametrize(
+    "model",
+    [
+        "us.anthropic.claude-opus-4-6",
+        "anthropic.claude-3-5-sonnet-20241022-v2:0",
+    ],
+)
+def test_legacy_thinking_forwarded_verbatim_when_model_accepts_it_converse(model):
+    """The 4.6 family and pre-adaptive models accept thinking={type: enabled}
+    natively, so the caller's budget_tokens cap must keep applying."""
+    config = AmazonConverseConfig()
+
+    optional_params = config.map_openai_params(
+        non_default_params={"thinking": {"type": "enabled", "budget_tokens": 4096}, "max_tokens": 8192},
+        optional_params={},
+        model=model,
+        drop_params=False,
+    )
+
+    assert optional_params["thinking"] == {"type": "enabled", "budget_tokens": 4096}
+    assert "output_config" not in optional_params
+
+
 def test_adaptive_thinking_dropped_when_max_tokens_too_small_converse():
     """When max_tokens can't fit even the minimum thinking budget, the raw
     adaptive block must be dropped entirely rather than translated, so the

tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py

@@ -422,6 +422,27 @@ def test_databricks_config_probes_capabilities_under_databricks_namespace():
     assert DatabricksConfig().custom_llm_provider == "databricks"
 
 
+@pytest.mark.parametrize(
+    "model, expected_thinking, expected_output_config",
+    [
+        ("databricks-claude-opus-4-8", {"type": "adaptive"}, {"effort": "high"}),
+        ("databricks-claude-opus-4-6", {"type": "enabled", "budget_tokens": 4096}, None),
+    ],
+    ids=["adaptive_only_upgrades_to_adaptive", "legacy_capable_forwards_verbatim"],
+)
+def test_map_openai_params_upgrades_legacy_thinking_on_adaptive_only_claude(
+    model, expected_thinking, expected_output_config
+):
+    mapped = DatabricksConfig().map_openai_params(
+        non_default_params={"thinking": {"type": "enabled", "budget_tokens": 4096}},
+        optional_params={},
+        model=model,
+        drop_params=False,
+    )
+    assert mapped["thinking"] == expected_thinking
+    assert mapped.get("output_config") == expected_output_config
+
+
 def _streaming_chunk(usage=None, choices=None):
     base = {
         "id": "chatcmpl-test",

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

@@ -752,3 +752,26 @@ def test_vertex_ai_fable_5_1_response_format_uses_native_output_format(local_mod
     assert "output_format" in result_params
     assert "tool_choice" not in result_params
     assert "tools" not in result_params
+
+
+def test_vertex_ai_anthropic_tool_based_response_format_still_upgrades_legacy_thinking(local_model_cost_map):
+    result_params = VertexAIAnthropicConfig().map_openai_params(
+        non_default_params={
+            "response_format": {
+                "type": "json_schema",
+                "json_schema": {
+                    "name": "test_schema",
+                    "schema": {"type": "object", "properties": {"result": {"type": "string"}}},
+                },
+            },
+            "thinking": {"type": "enabled", "budget_tokens": 4096},
+            "max_tokens": 8192,
+        },
+        optional_params={},
+        model="claude-opus-4-8",
+        drop_params=False,
+    )
+
+    assert "tools" in result_params
+    assert result_params["thinking"] == {"type": "adaptive"}
+    assert result_params["output_config"] == {"effort": "high"}