diff --git a/litellm/llms/anthropic/experimental_pass_through/adapters/handler.py b/litellm/llms/anthropic/experimental_pass_through/adapters/handler.py index 897ca3bf893..d16f5afb45c 100644 --- a/litellm/llms/anthropic/experimental_pass_through/adapters/handler.py +++ b/litellm/llms/anthropic/experimental_pass_through/adapters/handler.py @@ -106,6 +106,44 @@ class LiteLLMMessagesToCompletionTransformationHandler: updated_reasoning_effort["summary"] = effective_summary completion_kwargs["reasoning_effort"] = updated_reasoning_effort + @staticmethod + def _normalize_reasoning_effort( + completion_kwargs: Dict[str, Any], + ) -> None: + """ + Normalize reasoning_effort values based on target model capabilities. + + Handles both string ("max") and dict ({"effort": "max", "summary": ...}) + formats. Uses model registry to check supports_xhigh/supports_minimal. + """ + from litellm.llms.anthropic.experimental_pass_through.utils import ( + normalize_reasoning_effort_value, + ) + + reasoning_effort = completion_kwargs.get("reasoning_effort") + if reasoning_effort is None: + return + + model = cast(str, completion_kwargs.get("model", "")) + custom_llm_provider = completion_kwargs.get("custom_llm_provider") + + if isinstance(reasoning_effort, str): + normalized = normalize_reasoning_effort_value( + reasoning_effort, model=model, custom_llm_provider=custom_llm_provider + ) + if normalized != reasoning_effort: + completion_kwargs["reasoning_effort"] = normalized + elif isinstance(reasoning_effort, dict) and "effort" in reasoning_effort: + effort = reasoning_effort["effort"] + normalized = normalize_reasoning_effort_value( + effort, model=model, custom_llm_provider=custom_llm_provider + ) + if normalized != effort: + completion_kwargs["reasoning_effort"] = { + **reasoning_effort, + "effort": normalized, + } + @staticmethod def _prepare_completion_kwargs( *, @@ -163,6 +201,12 @@ class LiteLLMMessagesToCompletionTransformationHandler: if output_format: request_data["output_format"] = output_format + # Extract output_config from extra_kwargs so the translator can use it + # (e.g. output_config.effort for adaptive thinking → reasoning_effort) + extra_kwargs = extra_kwargs or {} + if "output_config" in extra_kwargs: + request_data["output_config"] = extra_kwargs["output_config"] + ( openai_request, tool_name_mapping, @@ -202,6 +246,14 @@ class LiteLLMMessagesToCompletionTransformationHandler: ): completion_kwargs[key] = value + # Normalize reasoning_effort based on model capabilities + # (e.g. "max" → "xhigh"/"high", "minimal" → "low" if unsupported) + # Must run BEFORE _route_openai_thinking, which prepends "responses/" + # to the model name and would break get_model_info() lookups. + LiteLLMMessagesToCompletionTransformationHandler._normalize_reasoning_effort( + completion_kwargs + ) + LiteLLMMessagesToCompletionTransformationHandler._route_openai_thinking_to_responses_api_if_needed( completion_kwargs, thinking=thinking, diff --git a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py index 072ae7c3bbe..e5d2b4ce782 100644 --- a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py +++ b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py @@ -317,6 +317,7 @@ class LiteLLMAnthropicMessagesAdapter: "tools", "thinking", "output_format", + "output_config", ] def _is_web_search_tool(self, tool: Dict[str, Any]) -> bool: @@ -694,6 +695,11 @@ class LiteLLMAnthropicMessagesAdapter: return "low" else: return "minimal" + elif thinking_type == "adaptive": + # Adaptive thinking: effort is controlled by output_config.effort, + # not budget_tokens. Return a default; caller should override with + # output_config.effort when available. + return "medium" return None @@ -1041,6 +1047,12 @@ class LiteLLMAnthropicMessagesAdapter: if not reasoning_effort: return + # For adaptive thinking, override with output_config.effort if available + if isinstance(thinking, dict) and thinking.get("type") == "adaptive": + output_config = anthropic_message_request.get("output_config") + if isinstance(output_config, dict) and output_config.get("effort"): + reasoning_effort = output_config["effort"] + summary = thinking.get("summary") if isinstance(thinking, dict) else None auto_summary = is_reasoning_auto_summary_enabled() if summary: diff --git a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/handler.py b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/handler.py index 198ebe1ff8c..5be16dcbf16 100644 --- a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/handler.py +++ b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/handler.py @@ -72,6 +72,23 @@ def _build_responses_kwargs( anthropic_request = AnthropicMessagesRequest(**request_data) # type: ignore[typeddict-item] responses_kwargs = _ADAPTER.translate_request(anthropic_request) + # Normalize reasoning effort based on model capabilities + # (e.g. "max" → "xhigh"/"high", "minimal" → "low" if unsupported) + reasoning = responses_kwargs.get("reasoning") + if isinstance(reasoning, dict) and "effort" in reasoning: + from litellm.llms.anthropic.experimental_pass_through.utils import ( + normalize_reasoning_effort_value, + ) + + effort = reasoning["effort"] + normalized = normalize_reasoning_effort_value( + effort, + model=model, + custom_llm_provider=(extra_kwargs or {}).get("custom_llm_provider"), + ) + if normalized != effort: + responses_kwargs["reasoning"] = {**reasoning, "effort": normalized} + if stream: responses_kwargs["stream"] = True diff --git a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py index 913470e7088..2badc2a3276 100644 --- a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py +++ b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py @@ -251,25 +251,41 @@ class LiteLLMAnthropicToResponsesAPIAdapter: @staticmethod def translate_thinking_to_reasoning( - thinking: Dict[str, Any] + thinking: Dict[str, Any], + output_config: Optional[Dict[str, Any]] = None, ) -> Optional[Dict[str, Any]]: """ Convert Anthropic thinking param to Responses API reasoning param. thinking.budget_tokens maps to reasoning effort: >= 10000 -> high, >= 5000 -> medium, >= 2000 -> low, < 2000 -> minimal + + For adaptive thinking, uses output_config.effort if available, + otherwise defaults to medium. """ - if not isinstance(thinking, dict) or thinking.get("type") != "enabled": + if not isinstance(thinking, dict): return None - budget = thinking.get("budget_tokens", 0) - if budget >= 10000: - effort = "high" - elif budget >= 5000: + + thinking_type = thinking.get("type") + + if thinking_type == "adaptive": + # Use output_config.effort if available effort = "medium" - elif budget >= 2000: - effort = "low" + if isinstance(output_config, dict) and output_config.get("effort"): + effort = output_config["effort"] + elif thinking_type == "enabled": + budget = thinking.get("budget_tokens", 0) + if budget >= 10000: + effort = "high" + elif budget >= 5000: + effort = "medium" + elif budget >= 2000: + effort = "low" + else: + effort = "minimal" else: - effort = "minimal" + return None + auto_summary = is_reasoning_auto_summary_enabled() result: Dict[str, Any] = {"effort": effort} summary = thinking.get("summary") @@ -346,7 +362,11 @@ class LiteLLMAnthropicToResponsesAPIAdapter: # thinking -> reasoning thinking = anthropic_request.get("thinking") if isinstance(thinking, dict): - reasoning = self.translate_thinking_to_reasoning(thinking) + output_config = anthropic_request.get("output_config") + reasoning = self.translate_thinking_to_reasoning( + thinking, + output_config=cast(Optional[Dict[str, Any]], output_config), + ) if reasoning: responses_kwargs["reasoning"] = reasoning diff --git a/litellm/llms/anthropic/experimental_pass_through/utils.py b/litellm/llms/anthropic/experimental_pass_through/utils.py index 6c1db6017b2..d975bee0bc2 100644 --- a/litellm/llms/anthropic/experimental_pass_through/utils.py +++ b/litellm/llms/anthropic/experimental_pass_through/utils.py @@ -1,4 +1,5 @@ import os +from typing import Optional import litellm @@ -9,3 +10,46 @@ def is_reasoning_auto_summary_enabled() -> bool: litellm.reasoning_auto_summary or os.getenv("LITELLM_REASONING_AUTO_SUMMARY", "false").lower() == "true" ) + + +def normalize_reasoning_effort_value( + effort: str, + model: str, + custom_llm_provider: Optional[str] = None, +) -> str: + """ + Normalize a reasoning effort value based on model capabilities. + + Degradation chains: + - "max" → max / xhigh / high + - "xhigh" → xhigh / high + - "minimal" → minimal / low + - other values pass through unchanged + """ + if effort not in ("max", "xhigh", "minimal"): + return effort + + from litellm.utils import get_model_info + + try: + model_info = get_model_info( + model=model, custom_llm_provider=custom_llm_provider + ) + except Exception: + model_info = {} + + if effort == "max": + if model_info.get("supports_max_reasoning_effort"): + return "max" + if model_info.get("supports_xhigh_reasoning_effort"): + return "xhigh" + return "high" + elif effort == "xhigh": + if model_info.get("supports_xhigh_reasoning_effort"): + return "xhigh" + return "high" + elif effort == "minimal": + if model_info.get("supports_minimal_reasoning_effort"): + return "minimal" + return "low" + return "medium" diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 04b68b8f4ec..05b59d45f99 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -1006,7 +1006,8 @@ "supports_vision": true, "tool_use_system_prompt_tokens": 346, "supports_native_structured_output": true, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "global.anthropic.claude-opus-4-6-v1": { "cache_creation_input_token_cost": 6.25e-06, @@ -1034,7 +1035,8 @@ "supports_vision": true, "tool_use_system_prompt_tokens": 346, "supports_native_structured_output": true, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "us.anthropic.claude-opus-4-6-v1": { "cache_creation_input_token_cost": 6.875e-06, @@ -1062,7 +1064,8 @@ "supports_vision": true, "tool_use_system_prompt_tokens": 346, "supports_native_structured_output": true, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "eu.anthropic.claude-opus-4-6-v1": { "cache_creation_input_token_cost": 6.875e-06, @@ -1090,7 +1093,8 @@ "supports_vision": true, "tool_use_system_prompt_tokens": 346, "supports_native_structured_output": true, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "au.anthropic.claude-opus-4-6-v1": { "cache_creation_input_token_cost": 6.875e-06, @@ -1118,7 +1122,8 @@ "supports_vision": true, "tool_use_system_prompt_tokens": 346, "supports_native_structured_output": true, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "anthropic.claude-opus-4-7": { "cache_creation_input_token_cost": 6.25e-06, @@ -1146,7 +1151,9 @@ "supports_vision": true, "supports_xhigh_reasoning_effort": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "global.anthropic.claude-opus-4-7": { "cache_creation_input_token_cost": 6.25e-06, @@ -1174,7 +1181,9 @@ "supports_vision": true, "supports_xhigh_reasoning_effort": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "us.anthropic.claude-opus-4-7": { "cache_creation_input_token_cost": 6.875e-06, @@ -1202,7 +1211,9 @@ "supports_vision": true, "supports_xhigh_reasoning_effort": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "eu.anthropic.claude-opus-4-7": { "cache_creation_input_token_cost": 6.875e-06, @@ -1230,7 +1241,9 @@ "supports_vision": true, "supports_xhigh_reasoning_effort": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "au.anthropic.claude-opus-4-7": { "cache_creation_input_token_cost": 6.875e-06, @@ -1258,7 +1271,9 @@ "supports_vision": true, "supports_xhigh_reasoning_effort": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "anthropic.claude-sonnet-4-6": { "cache_creation_input_token_cost": 3.75e-06, @@ -1285,7 +1300,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_minimal_reasoning_effort": true }, "global.anthropic.claude-sonnet-4-6": { "cache_creation_input_token_cost": 3.75e-06, @@ -1312,7 +1328,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_minimal_reasoning_effort": true }, "us.anthropic.claude-sonnet-4-6": { "cache_creation_input_token_cost": 4.125e-06, @@ -1339,7 +1356,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_minimal_reasoning_effort": true }, "eu.anthropic.claude-sonnet-4-6": { "cache_creation_input_token_cost": 4.125e-06, @@ -1366,7 +1384,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_minimal_reasoning_effort": true }, "au.anthropic.claude-sonnet-4-6": { "cache_creation_input_token_cost": 4.125e-06, @@ -1393,7 +1412,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_minimal_reasoning_effort": true }, "anthropic.claude-sonnet-4-20250514-v1:0": { "cache_creation_input_token_cost": 3.75e-06, @@ -1911,7 +1931,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 159, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "azure_ai/claude-opus-4-7": { "input_cost_per_token": 5e-06, @@ -1939,7 +1960,9 @@ "supports_tool_choice": true, "supports_vision": true, "supports_xhigh_reasoning_effort": true, - "tool_use_system_prompt_tokens": 159 + "tool_use_system_prompt_tokens": 159, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "azure_ai/claude-opus-4-1": { "cache_creation_input_token_cost": 1.875e-05, @@ -2003,7 +2026,8 @@ "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true, - "tool_use_system_prompt_tokens": 346 + "tool_use_system_prompt_tokens": 346, + "supports_minimal_reasoning_effort": true }, "azure/computer-use-preview": { "input_cost_per_token": 3e-06, @@ -8909,7 +8933,8 @@ "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true, - "tool_use_system_prompt_tokens": 346 + "tool_use_system_prompt_tokens": 346, + "supports_minimal_reasoning_effort": true }, "claude-sonnet-4-5-20250929-v1:0": { "cache_creation_input_token_cost": 3.75e-06, @@ -9103,7 +9128,8 @@ "us": 1.1, "fast": 6.0 }, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "claude-opus-4-6-20260205": { "cache_creation_input_token_cost": 6.25e-06, @@ -9135,7 +9161,8 @@ "us": 1.1, "fast": 6.0 }, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "claude-opus-4-7": { "cache_creation_input_token_cost": 6.25e-06, @@ -9167,7 +9194,9 @@ "provider_specific_entry": { "us": 1.1, "fast": 6.0 - } + }, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "claude-opus-4-7-20260416": { "cache_creation_input_token_cost": 6.25e-06, @@ -9199,7 +9228,9 @@ "provider_specific_entry": { "us": 1.1, "fast": 6.0 - } + }, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "claude-sonnet-4-20250514": { "deprecation_date": "2026-05-14", @@ -25052,7 +25083,8 @@ "supports_reasoning": true, "supports_tool_choice": true, "supports_vision": true, - "tool_use_system_prompt_tokens": 159 + "tool_use_system_prompt_tokens": 159, + "supports_minimal_reasoning_effort": true }, "openrouter/anthropic/claude-opus-4.5": { "cache_creation_input_token_cost": 6.25e-06, @@ -25090,7 +25122,8 @@ "supports_reasoning": true, "supports_tool_choice": true, "supports_vision": true, - "tool_use_system_prompt_tokens": 346 + "tool_use_system_prompt_tokens": 346, + "supports_minimal_reasoning_effort": true }, "openrouter/anthropic/claude-sonnet-4.5": { "input_cost_per_image": 0.0048, @@ -30118,7 +30151,8 @@ "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_minimal_reasoning_effort": true }, "vercel_ai_gateway/anthropic/claude-sonnet-4": { "cache_creation_input_token_cost": 3.75e-06, @@ -31345,7 +31379,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 346, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "vertex_ai/claude-opus-4-6@default": { "cache_creation_input_token_cost": 6.25e-06, @@ -31372,7 +31407,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 346, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "vertex_ai/claude-opus-4-7": { "cache_creation_input_token_cost": 6.25e-06, @@ -31399,7 +31435,9 @@ "supports_tool_choice": true, "supports_vision": true, "supports_xhigh_reasoning_effort": true, - "tool_use_system_prompt_tokens": 346 + "tool_use_system_prompt_tokens": 346, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "vertex_ai/claude-opus-4-7@default": { "cache_creation_input_token_cost": 6.25e-06, @@ -31426,7 +31464,9 @@ "supports_tool_choice": true, "supports_vision": true, "supports_xhigh_reasoning_effort": true, - "tool_use_system_prompt_tokens": 346 + "tool_use_system_prompt_tokens": 346, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "vertex_ai/claude-sonnet-4-5": { "cache_creation_input_token_cost": 3.75e-06, @@ -31478,7 +31518,8 @@ "search_context_size_high": 0.01, "search_context_size_low": 0.01, "search_context_size_medium": 0.01 - } + }, + "supports_minimal_reasoning_effort": true }, "vertex_ai/claude-sonnet-4-5@20250929": { "cache_creation_input_token_cost": 3.75e-06, @@ -38345,7 +38386,8 @@ "search_context_size_high": 0.01, "search_context_size_low": 0.01, "search_context_size_medium": 0.01 - } + }, + "supports_minimal_reasoning_effort": true }, "duckduckgo/search": { "litellm_provider": "duckduckgo", diff --git a/litellm/types/utils.py b/litellm/types/utils.py index e3058d106a6..c347956cba7 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -139,7 +139,9 @@ class ProviderSpecificModelInfo(TypedDict, total=False): supports_reasoning: Optional[bool] supports_url_context: Optional[bool] supports_none_reasoning_effort: Optional[bool] + supports_minimal_reasoning_effort: Optional[bool] supports_xhigh_reasoning_effort: Optional[bool] + supports_max_reasoning_effort: Optional[bool] class SearchContextCostPerQuery(TypedDict, total=False): diff --git a/litellm/utils.py b/litellm/utils.py index c4aee792972..7a9f62afa09 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -5893,9 +5893,15 @@ def _get_model_info_helper( # noqa: PLR0915 supports_none_reasoning_effort=_model_info.get( "supports_none_reasoning_effort", None ), + supports_minimal_reasoning_effort=_model_info.get( + "supports_minimal_reasoning_effort", None + ), supports_xhigh_reasoning_effort=_model_info.get( "supports_xhigh_reasoning_effort", None ), + supports_max_reasoning_effort=_model_info.get( + "supports_max_reasoning_effort", None + ), supports_computer_use=_model_info.get("supports_computer_use", None), search_context_cost_per_query=_model_info.get( "search_context_cost_per_query", None diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 386532f07a3..8a28235f985 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -1006,7 +1006,8 @@ "supports_vision": true, "tool_use_system_prompt_tokens": 346, "supports_native_structured_output": true, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "global.anthropic.claude-opus-4-6-v1": { "cache_creation_input_token_cost": 6.25e-06, @@ -1034,7 +1035,8 @@ "supports_vision": true, "tool_use_system_prompt_tokens": 346, "supports_native_structured_output": true, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "us.anthropic.claude-opus-4-6-v1": { "cache_creation_input_token_cost": 6.875e-06, @@ -1062,7 +1064,8 @@ "supports_vision": true, "tool_use_system_prompt_tokens": 346, "supports_native_structured_output": true, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "eu.anthropic.claude-opus-4-6-v1": { "cache_creation_input_token_cost": 6.875e-06, @@ -1090,7 +1093,8 @@ "supports_vision": true, "tool_use_system_prompt_tokens": 346, "supports_native_structured_output": true, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "au.anthropic.claude-opus-4-6-v1": { "cache_creation_input_token_cost": 6.875e-06, @@ -1118,7 +1122,8 @@ "supports_vision": true, "tool_use_system_prompt_tokens": 346, "supports_native_structured_output": true, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "anthropic.claude-opus-4-7": { "cache_creation_input_token_cost": 6.25e-06, @@ -1146,7 +1151,9 @@ "supports_vision": true, "supports_xhigh_reasoning_effort": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "anthropic.claude-mythos-preview": { "input_cost_per_token": 0, @@ -1188,7 +1195,9 @@ "supports_vision": true, "supports_xhigh_reasoning_effort": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "us.anthropic.claude-opus-4-7": { "cache_creation_input_token_cost": 6.875e-06, @@ -1216,7 +1225,9 @@ "supports_vision": true, "supports_xhigh_reasoning_effort": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "eu.anthropic.claude-opus-4-7": { "cache_creation_input_token_cost": 6.875e-06, @@ -1244,7 +1255,9 @@ "supports_vision": true, "supports_xhigh_reasoning_effort": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "au.anthropic.claude-opus-4-7": { "cache_creation_input_token_cost": 6.875e-06, @@ -1272,7 +1285,9 @@ "supports_vision": true, "supports_xhigh_reasoning_effort": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "anthropic.claude-sonnet-4-6": { "cache_creation_input_token_cost": 3.75e-06, @@ -1299,7 +1314,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_minimal_reasoning_effort": true }, "global.anthropic.claude-sonnet-4-6": { "cache_creation_input_token_cost": 3.75e-06, @@ -1326,7 +1342,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_minimal_reasoning_effort": true }, "us.anthropic.claude-sonnet-4-6": { "cache_creation_input_token_cost": 4.125e-06, @@ -1353,7 +1370,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_minimal_reasoning_effort": true }, "eu.anthropic.claude-sonnet-4-6": { "cache_creation_input_token_cost": 4.125e-06, @@ -1380,7 +1398,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_minimal_reasoning_effort": true }, "au.anthropic.claude-sonnet-4-6": { "cache_creation_input_token_cost": 4.125e-06, @@ -1407,7 +1426,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 346, - "supports_native_structured_output": true + "supports_native_structured_output": true, + "supports_minimal_reasoning_effort": true }, "anthropic.claude-sonnet-4-20250514-v1:0": { "cache_creation_input_token_cost": 3.75e-06, @@ -1925,7 +1945,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 159, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "azure_ai/claude-opus-4-7": { "input_cost_per_token": 5e-06, @@ -1953,7 +1974,9 @@ "supports_tool_choice": true, "supports_vision": true, "supports_xhigh_reasoning_effort": true, - "tool_use_system_prompt_tokens": 159 + "tool_use_system_prompt_tokens": 159, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "azure_ai/claude-opus-4-1": { "cache_creation_input_token_cost": 1.875e-05, @@ -2017,7 +2040,8 @@ "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true, - "tool_use_system_prompt_tokens": 346 + "tool_use_system_prompt_tokens": 346, + "supports_minimal_reasoning_effort": true }, "azure/computer-use-preview": { "input_cost_per_token": 3e-06, @@ -8923,7 +8947,8 @@ "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true, - "tool_use_system_prompt_tokens": 346 + "tool_use_system_prompt_tokens": 346, + "supports_minimal_reasoning_effort": true }, "claude-sonnet-4-5-20250929-v1:0": { "cache_creation_input_token_cost": 3.75e-06, @@ -9117,7 +9142,8 @@ "us": 1.1, "fast": 6.0 }, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "claude-opus-4-6-20260205": { "cache_creation_input_token_cost": 6.25e-06, @@ -9149,7 +9175,8 @@ "us": 1.1, "fast": 6.0 }, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "claude-opus-4-7": { "cache_creation_input_token_cost": 6.25e-06, @@ -9181,7 +9208,9 @@ "provider_specific_entry": { "us": 1.1, "fast": 6.0 - } + }, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "claude-opus-4-7-20260416": { "cache_creation_input_token_cost": 6.25e-06, @@ -9213,7 +9242,9 @@ "provider_specific_entry": { "us": 1.1, "fast": 6.0 - } + }, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "claude-sonnet-4-20250514": { "deprecation_date": "2026-05-14", @@ -25066,7 +25097,8 @@ "supports_reasoning": true, "supports_tool_choice": true, "supports_vision": true, - "tool_use_system_prompt_tokens": 159 + "tool_use_system_prompt_tokens": 159, + "supports_minimal_reasoning_effort": true }, "openrouter/anthropic/claude-opus-4.5": { "cache_creation_input_token_cost": 6.25e-06, @@ -25104,7 +25136,8 @@ "supports_reasoning": true, "supports_tool_choice": true, "supports_vision": true, - "tool_use_system_prompt_tokens": 346 + "tool_use_system_prompt_tokens": 346, + "supports_minimal_reasoning_effort": true }, "openrouter/anthropic/claude-sonnet-4.5": { "input_cost_per_image": 0.0048, @@ -30132,7 +30165,8 @@ "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_minimal_reasoning_effort": true }, "vercel_ai_gateway/anthropic/claude-sonnet-4": { "cache_creation_input_token_cost": 3.75e-06, @@ -31359,7 +31393,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 346, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "vertex_ai/claude-opus-4-6@default": { "cache_creation_input_token_cost": 6.25e-06, @@ -31386,7 +31421,8 @@ "supports_tool_choice": true, "supports_vision": true, "tool_use_system_prompt_tokens": 346, - "supports_max_reasoning_effort": true + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "vertex_ai/claude-opus-4-7": { "cache_creation_input_token_cost": 6.25e-06, @@ -31413,7 +31449,9 @@ "supports_tool_choice": true, "supports_vision": true, "supports_xhigh_reasoning_effort": true, - "tool_use_system_prompt_tokens": 346 + "tool_use_system_prompt_tokens": 346, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "vertex_ai/claude-opus-4-7@default": { "cache_creation_input_token_cost": 6.25e-06, @@ -31440,7 +31478,9 @@ "supports_tool_choice": true, "supports_vision": true, "supports_xhigh_reasoning_effort": true, - "tool_use_system_prompt_tokens": 346 + "tool_use_system_prompt_tokens": 346, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true }, "vertex_ai/claude-sonnet-4-5": { "cache_creation_input_token_cost": 3.75e-06, @@ -31492,7 +31532,8 @@ "search_context_size_high": 0.01, "search_context_size_low": 0.01, "search_context_size_medium": 0.01 - } + }, + "supports_minimal_reasoning_effort": true }, "vertex_ai/claude-sonnet-4-5@20250929": { "cache_creation_input_token_cost": 3.75e-06, @@ -38386,7 +38427,8 @@ "search_context_size_high": 0.01, "search_context_size_low": 0.01, "search_context_size_medium": 0.01 - } + }, + "supports_minimal_reasoning_effort": true }, "duckduckgo/search": { "litellm_provider": "duckduckgo", diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/test_reasoning_effort_fields.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/test_reasoning_effort_fields.py new file mode 100644 index 00000000000..d42d109f21b --- /dev/null +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/test_reasoning_effort_fields.py @@ -0,0 +1,287 @@ +""" +Tests for reasoning effort capability fields and normalize_reasoning_effort_value. + +Covers: +- Commit 1: get_model_info returns supports_minimal/supports_max fields +- Commit 2: Model registry entries have correct reasoning effort fields +- Commit 3: normalize_reasoning_effort_value degradation chains + adapter translation +""" + +import json +import os +from typing import Any, Dict, Optional +from unittest.mock import patch + +import pytest + +from litellm.llms.anthropic.experimental_pass_through.utils import ( + normalize_reasoning_effort_value, +) +from litellm.utils import get_model_info + + +def _load_model_registry() -> Dict[str, Any]: + """Load the root model_prices_and_context_window.json.""" + json_path = os.path.join( + os.path.dirname(__file__), + "../../../../../model_prices_and_context_window.json", + ) + with open(json_path) as f: + return json.load(f) + + +# --------------------------------------------------------------------------- +# Commit 1: get_model_info returns supports_minimal and supports_max fields +# --------------------------------------------------------------------------- + + +class TestGetModelInfoReasoningEffortFields: + """get_model_info should expose supports_minimal_reasoning_effort and + supports_max_reasoning_effort from the model registry.""" + + def test_opus_4_6_has_supports_minimal(self): + info = get_model_info("claude-opus-4-6") + assert "supports_minimal_reasoning_effort" in info + + def test_opus_4_6_has_supports_max(self): + info = get_model_info("claude-opus-4-6") + assert "supports_max_reasoning_effort" in info + + def test_opus_4_7_has_supports_minimal(self): + info = get_model_info("claude-opus-4-7") + assert "supports_minimal_reasoning_effort" in info + + def test_opus_4_7_has_supports_max(self): + info = get_model_info("claude-opus-4-7") + assert "supports_max_reasoning_effort" in info + + +# --------------------------------------------------------------------------- +# Commit 2: JSON registry has correct reasoning effort fields +# --------------------------------------------------------------------------- + + +class TestModelRegistryReasoningEffortFields: + """Verify specific models have the expected reasoning effort capability + values in the JSON registry file.""" + + @pytest.fixture(autouse=True) + def _load_registry(self): + self.registry = _load_model_registry() + + def test_opus_4_7_supports_max(self): + entry = self.registry["claude-opus-4-7"] + assert entry.get("supports_max_reasoning_effort") is True + + def test_opus_4_6_supports_max(self): + entry = self.registry["claude-opus-4-6"] + assert entry.get("supports_max_reasoning_effort") is True + + def test_opus_4_7_supports_minimal(self): + entry = self.registry["claude-opus-4-7"] + assert entry.get("supports_minimal_reasoning_effort") is True + + def test_opus_4_6_supports_minimal(self): + entry = self.registry["claude-opus-4-6"] + assert entry.get("supports_minimal_reasoning_effort") is True + + def test_sonnet_4_6_supports_minimal(self): + entry = self.registry["anthropic.claude-sonnet-4-6"] + assert entry.get("supports_minimal_reasoning_effort") is True + + def test_bedrock_opus_4_7_supports_max(self): + entry = self.registry["anthropic.claude-opus-4-7"] + assert entry.get("supports_max_reasoning_effort") is True + assert entry.get("supports_minimal_reasoning_effort") is True + + def test_vertex_opus_4_7_supports_max(self): + entry = self.registry["vertex_ai/claude-opus-4-7"] + assert entry.get("supports_max_reasoning_effort") is True + assert entry.get("supports_minimal_reasoning_effort") is True + + def test_vertex_opus_4_6_supports_max(self): + entry = self.registry["vertex_ai/claude-opus-4-6"] + assert entry.get("supports_max_reasoning_effort") is True + assert entry.get("supports_minimal_reasoning_effort") is True + + def test_azure_ai_opus_4_6_supports_minimal(self): + entry = self.registry["azure_ai/claude-opus-4-6"] + assert entry.get("supports_minimal_reasoning_effort") is True + + def test_azure_ai_opus_4_7_supports_max(self): + entry = self.registry["azure_ai/claude-opus-4-7"] + assert entry.get("supports_max_reasoning_effort") is True + assert entry.get("supports_minimal_reasoning_effort") is True + + +# --------------------------------------------------------------------------- +# Commit 3: normalize_reasoning_effort_value +# --------------------------------------------------------------------------- + + +def _mock_model_info(**flags): + """Return a mock model_info dict with given capability flags.""" + return flags + + +class TestNormalizeReasoningEffortValue: + """Test degradation chains for normalize_reasoning_effort_value.""" + + # --- "max" degradation chain --- + + def test_max_stays_max_when_supported(self): + with patch( + "litellm.utils.get_model_info", + return_value=_mock_model_info( + supports_max_reasoning_effort=True, + supports_xhigh_reasoning_effort=True, + ), + ): + assert normalize_reasoning_effort_value("max", model="test") == "max" + + def test_max_degrades_to_xhigh(self): + with patch( + "litellm.utils.get_model_info", + return_value=_mock_model_info( + supports_max_reasoning_effort=False, + supports_xhigh_reasoning_effort=True, + ), + ): + assert normalize_reasoning_effort_value("max", model="test") == "xhigh" + + def test_max_degrades_to_high(self): + with patch( + "litellm.utils.get_model_info", + return_value=_mock_model_info( + supports_max_reasoning_effort=False, + supports_xhigh_reasoning_effort=False, + ), + ): + assert normalize_reasoning_effort_value("max", model="test") == "high" + + # --- "xhigh" degradation chain --- + + def test_xhigh_stays_xhigh_when_supported(self): + with patch( + "litellm.utils.get_model_info", + return_value=_mock_model_info(supports_xhigh_reasoning_effort=True), + ): + assert normalize_reasoning_effort_value("xhigh", model="test") == "xhigh" + + def test_xhigh_degrades_to_high(self): + with patch( + "litellm.utils.get_model_info", + return_value=_mock_model_info(supports_xhigh_reasoning_effort=False), + ): + assert normalize_reasoning_effort_value("xhigh", model="test") == "high" + + # --- "minimal" degradation chain --- + + def test_minimal_stays_minimal_when_supported(self): + with patch( + "litellm.utils.get_model_info", + return_value=_mock_model_info(supports_minimal_reasoning_effort=True), + ): + assert ( + normalize_reasoning_effort_value("minimal", model="test") == "minimal" + ) + + def test_minimal_degrades_to_low(self): + with patch( + "litellm.utils.get_model_info", + return_value=_mock_model_info(supports_minimal_reasoning_effort=False), + ): + assert normalize_reasoning_effort_value("minimal", model="test") == "low" + + # --- passthrough values --- + + def test_high_passes_through(self): + assert normalize_reasoning_effort_value("high", model="test") == "high" + + def test_medium_passes_through(self): + assert normalize_reasoning_effort_value("medium", model="test") == "medium" + + def test_low_passes_through(self): + assert normalize_reasoning_effort_value("low", model="test") == "low" + + # --- exception fallback --- + + def test_exception_fallback_uses_empty_model_info(self): + """When get_model_info raises, treat model_info as {} (no capabilities).""" + with patch( + "litellm.utils.get_model_info", + side_effect=Exception("model not found"), + ): + # "max" with no capabilities -> "high" + assert normalize_reasoning_effort_value("max", model="unknown") == "high" + # "minimal" with no capabilities -> "low" + assert normalize_reasoning_effort_value("minimal", model="unknown") == "low" + + +# --------------------------------------------------------------------------- +# Commit 3: Adapter translation — adaptive thinking + output_config.effort +# --------------------------------------------------------------------------- + + +class TestAdapterAdaptiveThinking: + """Test that adaptive thinking type maps correctly through the adapters.""" + + def test_messages_adapter_adaptive_returns_medium_default(self): + """Adaptive thinking returns 'medium' as default reasoning_effort.""" + from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import ( + LiteLLMAnthropicMessagesAdapter, + ) + + adapter = LiteLLMAnthropicMessagesAdapter() + result = adapter.translate_anthropic_thinking_to_reasoning_effort( + {"type": "adaptive"} + ) + assert result == "medium" + + def test_messages_adapter_adaptive_overridden_by_output_config(self): + """For adaptive thinking, output_config.effort overrides reasoning_effort.""" + from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import ( + LiteLLMAnthropicMessagesAdapter, + ) + from litellm.types.llms.anthropic import AnthropicMessagesRequest + + adapter = LiteLLMAnthropicMessagesAdapter() + request = AnthropicMessagesRequest( + model="test-model", + messages=[{"role": "user", "content": "hello"}], + max_tokens=1024, + thinking={"type": "adaptive"}, + output_config={"effort": "high"}, + ) + openai_kwargs, _ = adapter.translate_anthropic_to_openai(request) + # reasoning_effort should be set (either as string or dict with effort) + re = openai_kwargs.get("reasoning_effort") + if isinstance(re, dict): + assert re["effort"] == "high" + else: + assert re == "high" + + def test_responses_adapter_adaptive_with_output_config(self): + """Responses adapter: adaptive thinking + output_config.effort.""" + from litellm.llms.anthropic.experimental_pass_through.responses_adapters.transformation import ( + LiteLLMAnthropicToResponsesAPIAdapter, + ) + + result = LiteLLMAnthropicToResponsesAPIAdapter.translate_thinking_to_reasoning( + thinking={"type": "adaptive"}, + output_config={"effort": "xhigh"}, + ) + assert result is not None + assert result["effort"] == "xhigh" + + def test_responses_adapter_adaptive_default_medium(self): + """Responses adapter: adaptive thinking without output_config defaults to medium.""" + from litellm.llms.anthropic.experimental_pass_through.responses_adapters.transformation import ( + LiteLLMAnthropicToResponsesAPIAdapter, + ) + + result = LiteLLMAnthropicToResponsesAPIAdapter.translate_thinking_to_reasoning( + thinking={"type": "adaptive"}, + ) + assert result is not None + assert result["effort"] == "medium"