From c6024f3d5f24a9fe7774cf419b3248b24117fcf1 Mon Sep 17 00:00:00 2001 From: Mateo Wang <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 3 Sep 2026 14:00:59 -0700 Subject: [PATCH] fix: treat gpt-6 names as the gpt-5 request family in OpenAI and Azure configs (#39631) Adapted for stable/1.98.x: requires_max_completion_tokens and its test are not on this line (added upstream by #36857), so that hunk is dropped. The Responses parametrize cases target a test class this line does not have, so the two gpt-6-astra cases are kept as a standalone test in the same file. (cherry picked from commit 025a3ca42fa81017e961815ef229bd04ff9e903a) --- .../llms/azure/chat/gpt_5_transformation.py | 22 ++-------------- .../llms/openai/chat/gpt_5_transformation.py | 26 ++++++++----------- .../llms/openai/responses/transformation.py | 3 ++- .../chat/test_azure_gpt5_transformation.py | 12 +++++++++ .../test_openai_responses_transformation.py | 10 +++++++ .../llms/openai/test_gpt5_transformation.py | 14 ++++++++++ .../llms/openai/test_is_model_gpt_5_model.py | 4 +++ tests/test_litellm/test_main.py | 13 ++++++++++ 8 files changed, 68 insertions(+), 36 deletions(-) diff --git a/litellm/llms/azure/chat/gpt_5_transformation.py b/litellm/llms/azure/chat/gpt_5_transformation.py index d7584083327..d72f907fc88 100644 --- a/litellm/llms/azure/chat/gpt_5_transformation.py +++ b/litellm/llms/azure/chat/gpt_5_transformation.py @@ -7,6 +7,7 @@ from litellm.exceptions import UnsupportedParamsError from litellm.llms.openai.chat.gpt_5_transformation import ( OpenAIGPT5Config, _get_effort_level, + is_gpt_reasoning_series_name, ) from litellm.types.llms.openai import AllMessageValues @@ -35,26 +36,7 @@ class AzureOpenAIGPT5Config(AzureOpenAIConfig, OpenAIGPT5Config): @classmethod def is_model_gpt_5_model(cls, model: str) -> bool: - """Check if the Azure model string refers to a gpt-5 variant. - - Accepts both explicit gpt-5 model names and the ``gpt5_series/`` prefix - used for manual routing. - """ - # The gpt-5-chat* family (gpt-5-chat, gpt-5-chat-latest, gpt-5-chat-2025-08-07, - # …) are regular chat models: they support temperature and tool_choice but NOT - # reasoning_effort. They must NOT be routed through the GPT-5 reasoning path. - # - # Versioned chat models such as gpt-5.3-chat and gpt-5.1-chat ARE reasoning - # models and must stay on the GPT-5 path. The distinguishing feature is that - # the gpt-5-chat family has a literal "-chat" immediately after "gpt-5" - # (i.e. "gpt-5-chat…"), while versioned chat models interpose a minor version - # number (i.e. "gpt-5.-chat"). - # - # Using a startswith("gpt-5-chat") prefix check on the normalized name (rather - # than a substring check) makes this boundary explicit and avoids any ambiguity - # if future model names coincidentally contain "gpt-5-chat" as an interior run. - _normalized: Final = model.split("/")[-1] # strip provider prefix, e.g. "azure/" - return ("gpt-5" in model and not _normalized.startswith("gpt-5-chat")) or "gpt5_series" in model + return is_gpt_reasoning_series_name(model) or "gpt5_series" in model def get_supported_openai_params(self, model: str) -> list[str]: """Get supported parameters for Azure OpenAI GPT-5 models. diff --git a/litellm/llms/openai/chat/gpt_5_transformation.py b/litellm/llms/openai/chat/gpt_5_transformation.py index ffa3de0d5c6..acb56be47e1 100644 --- a/litellm/llms/openai/chat/gpt_5_transformation.py +++ b/litellm/llms/openai/chat/gpt_5_transformation.py @@ -44,6 +44,14 @@ def _get_effort_level(value: str | dict | None) -> str | None: return None +GPT_REASONING_SERIES_MARKERS: Final = ("gpt-5", "gpt-6") + + +def is_gpt_reasoning_series_name(model: str) -> bool: + normalized: Final = model.split("/")[-1] + return any(marker in model for marker in GPT_REASONING_SERIES_MARKERS) and not normalized.startswith("gpt-5-chat") + + class OpenAIGPT5Config(OpenAIGPTConfig): """Configuration for gpt-5 models including GPT-5-Codex variants. @@ -56,21 +64,7 @@ class OpenAIGPT5Config(OpenAIGPTConfig): @classmethod def is_model_gpt_5_model(cls, model: str) -> bool: - # The gpt-5-chat* family (gpt-5-chat, gpt-5-chat-latest, gpt-5-chat-2025-08-07, - # …) are regular chat models: they support temperature and tool_choice but NOT - # reasoning_effort. They must NOT be routed through the GPT-5 reasoning path. - # - # Versioned chat models such as gpt-5.3-chat and gpt-5.1-chat ARE reasoning - # models and must stay on the GPT-5 path. The distinguishing feature is that - # the gpt-5-chat family has a literal "-chat" immediately after "gpt-5" - # (i.e. "gpt-5-chat…"), while versioned chat models interpose a minor version - # number (i.e. "gpt-5.-chat"). - # - # Using a startswith("gpt-5-chat") prefix check on the normalized name (rather - # than a substring check) makes this boundary explicit and avoids any ambiguity - # if future model names coincidentally contain "gpt-5-chat" as an interior run. - _normalized: Final = model.split("/")[-1] # strip provider prefix, e.g. "openai/" - return "gpt-5" in model and not _normalized.startswith("gpt-5-chat") + return is_gpt_reasoning_series_name(model) @classmethod def is_model_gpt_5_search_model(cls, model: str) -> bool: @@ -105,6 +99,8 @@ class OpenAIGPT5Config(OpenAIGPTConfig): def is_model_gpt_5_4_plus_model(cls, model: str) -> bool: """Check if the model is gpt-5.4 or newer (5.4, 5.5, 5.6, etc., including pro).""" model_name: Final = model.split("/")[-1] + if model_name.startswith("gpt-6"): + return True if not model_name.startswith("gpt-5."): return False try: diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py index b2a69564908..6c3c988ef54 100644 --- a/litellm/llms/openai/responses/transformation.py +++ b/litellm/llms/openai/responses/transformation.py @@ -12,6 +12,7 @@ from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response impo ) from litellm.litellm_core_utils.url_utils import encode_url_path_segment from litellm.llms.base_llm.responses.transformation import BaseResponsesAPIConfig +from litellm.llms.openai.chat.gpt_5_transformation import is_gpt_reasoning_series_name from litellm.secret_managers.main import get_secret_str from litellm.types.llms.openai import * from litellm.types.responses.main import * @@ -48,7 +49,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig): parts: Final = model.split("/") if len(parts) > 1 and parts[0] not in ("openai",): return False - return "gpt-5" in model and "gpt-5-chat" not in model + return is_gpt_reasoning_series_name(model) @staticmethod def _supports_reasoning_effort_none(model: str) -> bool: diff --git a/tests/test_litellm/llms/azure/chat/test_azure_gpt5_transformation.py b/tests/test_litellm/llms/azure/chat/test_azure_gpt5_transformation.py index 83562331b9a..aaa3f3f1b1c 100644 --- a/tests/test_litellm/llms/azure/chat/test_azure_gpt5_transformation.py +++ b/tests/test_litellm/llms/azure/chat/test_azure_gpt5_transformation.py @@ -299,3 +299,15 @@ def test_azure_gpt5_1_does_not_support_logprobs(config: AzureOpenAIGPT5Config): supported_params = config.get_supported_openai_params(model="gpt-5.1") assert "logprobs" not in supported_params assert "top_logprobs" not in supported_params + + +def test_azure_gpt_6_astra_takes_the_reasoning_series_request_shape(): + params = litellm.get_optional_params( + model="gpt-6-astra", + custom_llm_provider="azure", + max_tokens=100, + reasoning_effort="max", + ) + assert params["max_completion_tokens"] == 100 + assert "max_tokens" not in params + assert params["reasoning_effort"] == "max" diff --git a/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py b/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py index 151c51f1ca0..bab2dd567c3 100644 --- a/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py +++ b/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py @@ -1540,3 +1540,13 @@ class TestPhaseParameter: assert validated[0]["phase"] == "commentary" assert validated[1]["phase"] == "final_answer" assert "phase" not in validated[2] + + +@pytest.mark.parametrize("effort", [None, "low"]) +def test_gpt_6_astra_drops_temperature_on_the_responses_path(effort): + mapped = OpenAIResponsesAPIConfig().map_openai_params( + response_api_optional_params={"temperature": 0, **({"reasoning": {"effort": effort}} if effort else {})}, + model="gpt-6-astra", + drop_params=True, + ) + assert "temperature" not in mapped diff --git a/tests/test_litellm/llms/openai/test_gpt5_transformation.py b/tests/test_litellm/llms/openai/test_gpt5_transformation.py index d279b119efe..d8e8feac1e3 100644 --- a/tests/test_litellm/llms/openai/test_gpt5_transformation.py +++ b/tests/test_litellm/llms/openai/test_gpt5_transformation.py @@ -1309,3 +1309,17 @@ def test_responses_gpt54_allow_temperature_effort_none( drop_params=False, ) assert params["temperature"] == 0.7 + + +def test_gpt_6_astra_takes_the_reasoning_series_request_shape(): + params = litellm.get_optional_params( + model="gpt-6-astra", + custom_llm_provider="openai", + max_tokens=100, + reasoning_effort="max", + verbosity="low", + ) + assert params["max_completion_tokens"] == 100 + assert "max_tokens" not in params + assert params["reasoning_effort"] == "max" + assert params["verbosity"] == "low" diff --git a/tests/test_litellm/llms/openai/test_is_model_gpt_5_model.py b/tests/test_litellm/llms/openai/test_is_model_gpt_5_model.py index 1095819c98c..107a1afb2c6 100644 --- a/tests/test_litellm/llms/openai/test_is_model_gpt_5_model.py +++ b/tests/test_litellm/llms/openai/test_is_model_gpt_5_model.py @@ -41,6 +41,8 @@ from litellm.llms.azure.chat.gpt_5_transformation import AzureOpenAIGPT5Config # Models that MUST be classified as GPT-5 (routed through GPT-5 reasoning path) GPT5_MODELS = [ + "gpt-6-astra", + "openai/gpt-6-astra", "gpt-5", "gpt-5.1", "gpt-5.2", @@ -120,6 +122,8 @@ class TestOpenAIGPT5ConfigIsModelGpt5Model: # /v1/responses bridge (when reasoning_effort is set and tools are passed) on # is_model_gpt_5_4_plus_model, so the gpt-5.6 family must land on the True side. GPT5_4_PLUS_MODELS = [ + "gpt-6-astra", + "openai/gpt-6-astra", "gpt-5.4", "gpt-5.5", "gpt-5.5-pro", diff --git a/tests/test_litellm/test_main.py b/tests/test_litellm/test_main.py index 58373df024c..656b12ce044 100644 --- a/tests/test_litellm/test_main.py +++ b/tests/test_litellm/test_main.py @@ -778,6 +778,19 @@ def test_responses_api_bridge_check_gpt_5_4_tools_plus_reasoning_routes_to_respo assert model_info.get("mode") == "responses" +def test_responses_api_bridge_check_gpt_6_astra_tools_with_default_reasoning_routes_to_responses(): + from litellm.main import responses_api_bridge_check + + model_info, model = responses_api_bridge_check( + model="gpt-6-astra", + custom_llm_provider="openai", + tools=[{"type": "function", "function": {"name": "get_capital"}}], + ) + + assert model == "gpt-6-astra" + assert model_info.get("mode") == "responses" + + def test_responses_api_bridge_check_gpt_5_5_tools_plus_reasoning_routes_to_responses(): """gpt-5.5+ with both tools and reasoning_effort should route to Responses API.""" from litellm.main import responses_api_bridge_check