Fix: Azure OpenAI Responses API incorrectly blocking temperature parameter for GPT-5

[xingyaoww]

Problem

Azure OpenAI Responses API was incorrectly blocking the temperature parameter for GPT-5 models with the following error:

litellm.exceptions.UnsupportedParamsError: LlmProviders.AZURE does not support parameters: {'temperature': 1.0}, for model=gpt-5

However, according to both OpenAI documentation and Azure AI documentation, the Responses API does support the temperature parameter for non-O-series models.

Source issue: fix OpenHands/software-agent-sdk#986

Root Cause

The bug occurred because LiteLLM was using supports_reasoning(model) as a proxy to detect O-series models (o1, o3, o4). This caused GPT-5 to be incorrectly classified:

1. In litellm/utils.py (line 7400): Used supports_reasoning(model) to determine if a model should use AzureOpenAIOSeriesResponsesAPIConfig
2. In litellm/llms/azure/responses/o_series_transformation.py (line 93): The is_o_series_model() method also used supports_reasoning(model)
3. The Issue: GPT-5 has "supports_reasoning": true in model_prices_and_context_window.json, but GPT-5 is NOT an O-series model
4. Result: GPT-5 incorrectly used AzureOpenAIOSeriesResponsesAPIConfig which blocks temperature, instead of AzureOpenAIResponsesAPIConfig which allows it

The Key Distinction

• O-series models are specific model families: o1, o3, o4 (e.g., o1-preview, o3-mini)
  • These models have parameter restrictions (no temperature, etc.)
• Models that support reasoning is a broader category that includes:
  • O-series models (o1, o3, o4)
  • GPT-5 (supports reasoning but is NOT O-series)

Solution

Updated the O-series detection logic to check for specific model name patterns instead of using the supports_reasoning flag:

Changes Made

1. litellm/llms/azure/responses/o_series_transformation.py:

   • Updated is_o_series_model() to check for explicit O-series model names: "o1", "o3", "o4", or "o_series" in model name
   • Removed dependency on supports_reasoning()
   • Removed unused supports_reasoning import

2. litellm/utils.py:

   • Updated ProviderConfigManager.get_provider_responses_api_config() to use explicit O-series model name checks
   • No longer relies on supports_reasoning() for determining config selection

3. tests/test_litellm/llms/azure/response/test_azure_transformation.py:

   • Enhanced test_o_series_model_detection() to include GPT-5 test cases
   • Enhanced test_provider_config_manager_o_series_selection() to verify:
     • GPT-5 gets AzureOpenAIResponsesAPIConfig (not O-series config)
     • GPT-5 supports the temperature parameter
     • O1/O3 models still correctly get AzureOpenAIOSeriesResponsesAPIConfig

Verification

Before Fix

config = ProviderConfigManager.get_provider_responses_api_config(
    provider=litellm.LlmProviders.AZURE, model="gpt-5"
)
# Returns: AzureOpenAIOSeriesResponsesAPIConfig (WRONG!)
# Temperature supported: False (BUG!)

After Fix

config = ProviderConfigManager.get_provider_responses_api_config(
    provider=litellm.LlmProviders.AZURE, model="gpt-5"
)
# Returns: AzureOpenAIResponsesAPIConfig (CORRECT!)
# Temperature supported: True (FIXED!)

Test Results

Model: gpt-5          (Regular model with reasoning support)
  Config: AzureOpenAIResponsesAPIConfig
  Supports temperature: True ✓

Model: gpt-4o         (Regular model)
  Config: AzureOpenAIResponsesAPIConfig
  Supports temperature: True ✓

Model: o1-preview     (O-series model)
  Config: AzureOpenAIOSeriesResponsesAPIConfig
  Supports temperature: False ✓

Model: o3-mini        (O-series model)
  Config: AzureOpenAIOSeriesResponsesAPIConfig
  Supports temperature: False ✓

All 15 tests in test_azure_transformation.py pass ✅

Impact

• ✅ GPT-5 now correctly supports the temperature parameter on Azure Responses API
• ✅ O-series models (o1, o3, o4) continue to work correctly with appropriate restrictions
• ✅ All existing functionality remains intact
• ✅ No breaking changes

Files Changed

• litellm/llms/azure/responses/o_series_transformation.py (+7, -6)
• litellm/utils.py (+4, -2)
• tests/test_litellm/llms/azure/response/test_azure_transformation.py (+38, -0)

Total: 3 files changed, 48 insertions(+), 9 deletions(-)

@xingyaoww can click here to continue refining the PR

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Preview	Comments	Updated (UTC)
litellm	Ready	Preview	Comment	Nov 4, 2025 2:30pm

[krrishdholakia]

this change seems wrong @xingyaoww , shouldn't it be to make sure gpt-5 is not routed into the o-series logic?

[xingyaoww]

@krrishdholakia maybe this is the correct fix: #16246