From e7454e52e6b15c9d7cce043abb84d43eecb1f0d7 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Tue, 15 Sep 2026 07:09:05 +0000 Subject: [PATCH 1/2] fix(gemini): map minimal thinking to low for Gemini 3.7 and 3.8 Flash Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../vertex_and_google_ai_studio_gemini.py | 35 ++++----- ...test_vertex_and_google_ai_studio_gemini.py | 78 +++++++++++++++++++ 2 files changed, 95 insertions(+), 18 deletions(-) diff --git a/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py b/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py index d113b2b4f6b..d719d53e19f 100644 --- a/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py +++ b/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py @@ -110,6 +110,7 @@ else: SUPPORTED_REASONING_EFFORTS: Final = ("minimal", "low", "medium", "high", "none", "disable") +GEMINI_FLASH_MODELS_WITHOUT_MINIMAL_THINKING: Final = ("gemini-3.7-flash", "gemini-3.8-flash") def _unsupported_reasoning_effort(reasoning_effort: str) -> UnsupportedParamsError: @@ -860,6 +861,12 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig): else: raise _unsupported_reasoning_effort(reasoning_effort) + @staticmethod + def _supports_minimal_thinking_level(model: str) -> bool: + lowered: Final = model.lower() + is_gemini3flash: Final = "gemini-3" in lowered and "flash" in lowered + return is_gemini3flash and not any(m in lowered for m in GEMINI_FLASH_MODELS_WITHOUT_MINIMAL_THINKING) + @staticmethod def _map_reasoning_effort_to_thinking_level( reasoning_effort: str, @@ -874,13 +881,11 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig): Returns: GeminiThinkingConfig with thinkingLevel and includeThoughts """ - # Check if this is gemini-3-flash which supports MINIMAL thinking level - # Covers gemini-3-flash, gemini-3-flash-preview, gemini-3.1-flash, gemini-3.1-flash-lite-preview, - # gemini-3.5-flash, and any future 3.x-flash variants. is_gemini3flash: Final = model and ("flash" in model.lower() and "gemini-3" in model.lower()) + supports_minimal: Final = bool(model) and VertexGeminiConfig._supports_minimal_thinking_level(model) is_gemini31pro: Final = model and ("gemini-3.1-pro-preview" in model.lower()) if reasoning_effort == "minimal": - if is_gemini3flash: + if supports_minimal: return {"thinkingLevel": "minimal", "includeThoughts": True} else: return {"thinkingLevel": "low", "includeThoughts": True} @@ -893,18 +898,11 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig): return {"thinkingLevel": "high", "includeThoughts": True} elif reasoning_effort == "high": return {"thinkingLevel": "high", "includeThoughts": True} - elif reasoning_effort == "disable": - # Gemini 3 cannot fully disable thinking, so we use "minimal" for gemini-3-flash-preview, "low" for others - if is_gemini3flash: - return {"thinkingLevel": "minimal", "includeThoughts": False} - else: - return {"thinkingLevel": "low", "includeThoughts": False} - elif reasoning_effort == "none": - # For gemini-3-flash-preview, use "minimal" instead of "low" - if is_gemini3flash: - return {"thinkingLevel": "minimal", "includeThoughts": False} - else: - return {"thinkingLevel": "low", "includeThoughts": False} + elif reasoning_effort in ("disable", "none"): + return { + "thinkingLevel": "minimal" if supports_minimal else "low", + "includeThoughts": False, + } else: raise _unsupported_reasoning_effort(reasoning_effort) @@ -971,8 +969,9 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig): params["includeThoughts"] = True # Follow provider defaults unless explicitly opted into legacy behavior. if litellm.enable_gemini_default_thinking_level_low is True: - is_gemini3flash: Final = "gemini-3" in model.lower() and "flash" in model.lower() - params["thinkingLevel"] = "minimal" if is_gemini3flash else "low" + params["thinkingLevel"] = ( + "minimal" if VertexGeminiConfig._supports_minimal_thinking_level(model) else "low" + ) else: # Thinking disabled params["includeThoughts"] = False diff --git a/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py b/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py index 101f6e6fa5d..a1c31689d09 100644 --- a/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py +++ b/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py @@ -2678,6 +2678,84 @@ def test_reasoning_effort_maps_to_thinking_level_gemini_3(): assert result["thinkingConfig"]["includeThoughts"] is False +@pytest.mark.parametrize( + "model", + [ + "gemini-3.7-flash", + "vertex_ai/gemini-3.8-flash", + "gemini-3.8-flash-preview", + ], +) +@pytest.mark.parametrize( + ("reasoning_effort", "include_thoughts"), + [("minimal", True), ("none", False), ("disable", False)], +) +def test_gemini_37_38_flash_floor_minimal_thinking_level( + model, reasoning_effort, include_thoughts +): + result = VertexGeminiConfig._map_reasoning_effort_to_thinking_level( + reasoning_effort, model + ) + + assert result["thinkingLevel"] == "low" + assert result["includeThoughts"] is include_thoughts + + +@pytest.mark.parametrize( + ("model", "reasoning_effort", "expected_level", "include_thoughts"), + [ + ("gemini-3-flash-preview", "minimal", "minimal", True), + ("gemini-3-flash-preview", "none", "minimal", False), + ("gemini-3-flash-preview", "disable", "minimal", False), + ("gemini-3.6-flash", "minimal", "minimal", True), + ("gemini-3.6-flash", "none", "minimal", False), + ("gemini-3.6-flash", "disable", "minimal", False), + ("gemini-3.5-flash", "minimal", "minimal", True), + ("gemini-3.5-flash", "none", "minimal", False), + ("gemini-3.5-flash", "disable", "minimal", False), + ("gemini-3.8-flash", "medium", "medium", True), + ], +) +def test_gemini_flash_minimal_thinking_support( + model, reasoning_effort, expected_level, include_thoughts +): + result = VertexGeminiConfig._map_reasoning_effort_to_thinking_level( + reasoning_effort, model + ) + + assert result["thinkingLevel"] == expected_level + assert result["includeThoughts"] is include_thoughts + + +def test_gemini_38_flash_feature_flag_uses_low_thinking_level(monkeypatch): + monkeypatch.setattr(litellm, "enable_gemini_default_thinking_level_low", True) + thinking_param = {"type": "enabled", "budget_tokens": 1024} + + result_38 = VertexGeminiConfig._map_thinking_param( + thinking_param, model="gemini-3.8-flash" + ) + result_36 = VertexGeminiConfig._map_thinking_param( + thinking_param, model="gemini-3.6-flash" + ) + + assert result_38["thinkingLevel"] == "low" + assert result_36["thinkingLevel"] == "minimal" + + +def test_gemini_38_flash_public_reasoning_effort_none_uses_low(): + result = VertexGeminiConfig().map_openai_params( + non_default_params={"reasoning_effort": "none"}, + optional_params={}, + model="gemini-3.8-flash", + drop_params=False, + ) + + assert result["thinkingConfig"] == { + "thinkingLevel": "low", + "includeThoughts": False, + } + + def test_reasoning_effort_dict_format_gemini_3(): """ Test that reasoning_effort works when passed as dict format from OpenAI Agents SDK. From a1ad95dbbd5dc9003598278c610b401f797d6403 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 16 Sep 2026 16:13:41 -0700 Subject: [PATCH 2/2] fix(gemini): read the minimal thinking floor from the cost map and cover the /v1/messages bridge --- .../llms/openai/chat/gpt_5_transformation.py | 4 +- .../vertex_and_google_ai_studio_gemini.py | 6 ++- ...odel_prices_and_context_window_backup.json | 6 +++ litellm/utils.py | 6 +-- model_prices_and_context_window.json | 6 +++ .../llms/openai/test_gpt5_transformation.py | 16 +++---- ...test_vertex_and_google_ai_studio_gemini.py | 47 +++++++++++++++++-- 7 files changed, 71 insertions(+), 20 deletions(-) diff --git a/litellm/llms/openai/chat/gpt_5_transformation.py b/litellm/llms/openai/chat/gpt_5_transformation.py index b02f953425d..1b93df95341 100644 --- a/litellm/llms/openai/chat/gpt_5_transformation.py +++ b/litellm/llms/openai/chat/gpt_5_transformation.py @@ -4,9 +4,9 @@ from typing import Final import litellm from litellm.utils import ( - _is_explicitly_disabled_factory, _supports_factory, declared_value_factory, + is_explicitly_disabled_factory, ) from .gpt_transformation import OpenAIGPTConfig @@ -192,7 +192,7 @@ class OpenAIGPT5Config(OpenAIGPTConfig): Use this for opt-out checks where unknown models should be allowed through. """ - return _is_explicitly_disabled_factory( + return is_explicitly_disabled_factory( model=cls._model_map_lookup_name(model), custom_llm_provider=None, key=f"supports_{level}_reasoning_effort", diff --git a/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py b/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py index d719d53e19f..7d616c37ec1 100644 --- a/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py +++ b/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py @@ -79,6 +79,7 @@ from litellm.utils import ( CustomStreamWrapper, ModelResponse, is_base64_encoded, + is_explicitly_disabled_factory, supports_reasoning, ) @@ -110,7 +111,6 @@ else: SUPPORTED_REASONING_EFFORTS: Final = ("minimal", "low", "medium", "high", "none", "disable") -GEMINI_FLASH_MODELS_WITHOUT_MINIMAL_THINKING: Final = ("gemini-3.7-flash", "gemini-3.8-flash") def _unsupported_reasoning_effort(reasoning_effort: str) -> UnsupportedParamsError: @@ -865,7 +865,9 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig): def _supports_minimal_thinking_level(model: str) -> bool: lowered: Final = model.lower() is_gemini3flash: Final = "gemini-3" in lowered and "flash" in lowered - return is_gemini3flash and not any(m in lowered for m in GEMINI_FLASH_MODELS_WITHOUT_MINIMAL_THINKING) + return is_gemini3flash and not is_explicitly_disabled_factory( + model=model, custom_llm_provider=None, key="supports_minimal_reasoning_effort" + ) @staticmethod def _map_reasoning_effort_to_thinking_level( diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 9f91cf82f41..43399d5af53 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -25159,6 +25159,7 @@ "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_minimal_reasoning_effort": false, "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, @@ -25216,6 +25217,7 @@ "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_minimal_reasoning_effort": false, "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, @@ -27081,6 +27083,7 @@ "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_minimal_reasoning_effort": false, "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, @@ -27140,6 +27143,7 @@ "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_minimal_reasoning_effort": false, "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, @@ -27555,6 +27559,7 @@ "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_minimal_reasoning_effort": false, "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, @@ -27612,6 +27617,7 @@ "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_minimal_reasoning_effort": false, "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, diff --git a/litellm/utils.py b/litellm/utils.py index 18df5e2abf7..b6f4e85a702 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -2675,7 +2675,7 @@ def declared_value_factory(model: str, custom_llm_provider: str | None, key: str """Return a string value the model map declares for *key*, or ``None`` when it says nothing. The string-valued sibling of :func:`_supports_factory` and - :func:`_is_explicitly_disabled_factory`, public where those two are not because it is read + :func:`is_explicitly_disabled_factory`, public like the latter because both are read from the provider configs rather than from this module, sharing their ``get_llm_provider`` -> ``_get_model_info_helper`` chain and their unprefixed-twin fallback (#20885), so a provider-prefixed entry that omits the key still answers @@ -2711,7 +2711,7 @@ def declared_value_factory(model: str, custom_llm_provider: str | None, key: str return None -def _is_explicitly_disabled_factory(model: str, custom_llm_provider: str | None, key: str) -> bool: +def is_explicitly_disabled_factory(model: str, custom_llm_provider: str | None, key: str) -> bool: """Return True only when the model map explicitly sets *key* to ``False``. This is the opt-out mirror of :func:`_supports_factory`. Where @@ -2830,7 +2830,7 @@ def is_vision_explicitly_disabled(model: str, custom_llm_provider: str | None = The opt-out mirror of :func:`supports_vision`: a missing declaration reads as not disabled, so unknown or newly added models stay eligible for image routing. """ - return _is_explicitly_disabled_factory(model, custom_llm_provider, "supports_vision") + return is_explicitly_disabled_factory(model, custom_llm_provider, "supports_vision") def supports_vision(model: str, custom_llm_provider: str | None = None) -> bool: diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 9f91cf82f41..43399d5af53 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -25159,6 +25159,7 @@ "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_minimal_reasoning_effort": false, "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, @@ -25216,6 +25217,7 @@ "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_minimal_reasoning_effort": false, "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, @@ -27081,6 +27083,7 @@ "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_minimal_reasoning_effort": false, "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, @@ -27140,6 +27143,7 @@ "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_minimal_reasoning_effort": false, "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, @@ -27555,6 +27559,7 @@ "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_minimal_reasoning_effort": false, "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, @@ -27612,6 +27617,7 @@ "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_minimal_reasoning_effort": false, "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, diff --git a/tests/test_litellm/llms/openai/test_gpt5_transformation.py b/tests/test_litellm/llms/openai/test_gpt5_transformation.py index b538fad71a2..ba51209e0d5 100644 --- a/tests/test_litellm/llms/openai/test_gpt5_transformation.py +++ b/tests/test_litellm/llms/openai/test_gpt5_transformation.py @@ -8,7 +8,7 @@ from litellm.litellm_core_utils.get_model_cost_map import get_model_cost_map from litellm.llms.openai.chat.gpt_5_transformation import OpenAIGPT5Config from litellm.llms.openai.openai import OpenAIConfig from litellm.utils import ( - _is_explicitly_disabled_factory, + is_explicitly_disabled_factory, peek_reasoning_summary_aliases, strip_reasoning_summary_aliases_from_optional_params, ) @@ -524,19 +524,19 @@ def test_gpt5_minimal_explicitly_disabled_check(gpt5_config: OpenAIGPT5Config): def test_is_explicitly_disabled_factory_minimal(): - """_is_explicitly_disabled_factory returns True only for explicit False entries. + """is_explicitly_disabled_factory returns True only for explicit False entries. Verifies the shared helper used by _is_reasoning_effort_level_explicitly_disabled directly — so future changes to the helper are caught without going through the method wrapper. """ key = "supports_minimal_reasoning_effort" - assert _is_explicitly_disabled_factory("gpt-5.4-mini", None, key) - assert _is_explicitly_disabled_factory("gpt-5.4-nano", None, key) - assert _is_explicitly_disabled_factory("openai/gpt-5.4-mini", None, key) - assert _is_explicitly_disabled_factory("gpt-5.4", None, key) - assert _is_explicitly_disabled_factory("gpt-5.4-pro", None, key) - assert not _is_explicitly_disabled_factory("gpt-5.4-turbo-preview", None, key) + assert is_explicitly_disabled_factory("gpt-5.4-mini", None, key) + assert is_explicitly_disabled_factory("gpt-5.4-nano", None, key) + assert is_explicitly_disabled_factory("openai/gpt-5.4-mini", None, key) + assert is_explicitly_disabled_factory("gpt-5.4", None, key) + assert is_explicitly_disabled_factory("gpt-5.4-pro", None, key) + assert not is_explicitly_disabled_factory("gpt-5.4-turbo-preview", None, key) def test_gpt5_unknown_model_passes_through_minimal(config: OpenAIConfig): diff --git a/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py b/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py index a1c31689d09..b36c4e6c205 100644 --- a/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py +++ b/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py @@ -5,11 +5,14 @@ from copy import deepcopy from typing import Final, List, cast from unittest.mock import MagicMock, patch +import httpx import pytest from pydantic import BaseModel import litellm from litellm import ModelResponse, completion +from litellm.llms.anthropic.experimental_pass_through.messages import handler as anthropic_messages_handler +from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler from litellm.llms.gemini.chat.transformation import GoogleAIStudioGeminiConfig from litellm.llms.vertex_ai.common_utils import VertexAIError from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import ( @@ -2683,7 +2686,7 @@ def test_reasoning_effort_maps_to_thinking_level_gemini_3(): [ "gemini-3.7-flash", "vertex_ai/gemini-3.8-flash", - "gemini-3.8-flash-preview", + "gemini/gemini-3.8-flash", ], ) @pytest.mark.parametrize( @@ -2691,7 +2694,7 @@ def test_reasoning_effort_maps_to_thinking_level_gemini_3(): [("minimal", True), ("none", False), ("disable", False)], ) def test_gemini_37_38_flash_floor_minimal_thinking_level( - model, reasoning_effort, include_thoughts + local_model_cost_map, model, reasoning_effort, include_thoughts ): result = VertexGeminiConfig._map_reasoning_effort_to_thinking_level( reasoning_effort, model @@ -2717,7 +2720,7 @@ def test_gemini_37_38_flash_floor_minimal_thinking_level( ], ) def test_gemini_flash_minimal_thinking_support( - model, reasoning_effort, expected_level, include_thoughts + local_model_cost_map, model, reasoning_effort, expected_level, include_thoughts ): result = VertexGeminiConfig._map_reasoning_effort_to_thinking_level( reasoning_effort, model @@ -2727,7 +2730,7 @@ def test_gemini_flash_minimal_thinking_support( assert result["includeThoughts"] is include_thoughts -def test_gemini_38_flash_feature_flag_uses_low_thinking_level(monkeypatch): +def test_gemini_38_flash_feature_flag_uses_low_thinking_level(local_model_cost_map, monkeypatch): monkeypatch.setattr(litellm, "enable_gemini_default_thinking_level_low", True) thinking_param = {"type": "enabled", "budget_tokens": 1024} @@ -2742,7 +2745,7 @@ def test_gemini_38_flash_feature_flag_uses_low_thinking_level(monkeypatch): assert result_36["thinkingLevel"] == "minimal" -def test_gemini_38_flash_public_reasoning_effort_none_uses_low(): +def test_gemini_38_flash_public_reasoning_effort_none_uses_low(local_model_cost_map): result = VertexGeminiConfig().map_openai_params( non_default_params={"reasoning_effort": "none"}, optional_params={}, @@ -2756,6 +2759,40 @@ def test_gemini_38_flash_public_reasoning_effort_none_uses_low(): } +@pytest.mark.asyncio +async def test_gemini_38_flash_messages_bridge_thinking_disabled_sends_low_thinking_level(local_model_cost_map): + captured: dict[str, dict] = {} + + def upstream(request: httpx.Request) -> httpx.Response: + captured["body"] = json.loads(request.content) + return httpx.Response( + 200, + json={ + "candidates": [{"content": {"parts": [{"text": "hi"}], "role": "model"}, "finishReason": "STOP"}], + "usageMetadata": {"promptTokenCount": 1, "candidatesTokenCount": 1, "totalTokenCount": 2}, + }, + request=request, + ) + + client = AsyncHTTPHandler() + client.client = httpx.AsyncClient(transport=httpx.MockTransport(upstream)) + + await anthropic_messages_handler.anthropic_messages( + max_tokens=16, + messages=[{"role": "user", "content": "hi"}], + model="gemini/gemini-3.8-flash", + custom_llm_provider="gemini", + thinking={"type": "disabled"}, + api_key="fake-gemini-key", + client=client, + ) + + assert captured["body"]["generationConfig"]["thinkingConfig"] == { + "thinkingLevel": "low", + "includeThoughts": False, + } + + def test_reasoning_effort_dict_format_gemini_3(): """ Test that reasoning_effort works when passed as dict format from OpenAI Agents SDK.