GPT-5 fail gracefully if temperature parameter is provided

[gusfraser]

Description

If temperature is sent as a parameter e.g.

agent = PydanticAgent(
                model="openai:gpt-5-mini",
                instructions="You are an expert at evaluating whether conversation goals have been achieved. Be precise and analytical.",
                model_settings={"temperature": 0.7, "max_tokens": 200}  
            )

This fails with gpt-5 models as temperature is no longer a parameter.

PydanticAI generation failed: status_code: 400, model_name: gpt-5-mini, body: {'message': "Unsupported value: 'temperature' does not support 0.7 with this model. Only the default (1) value is supported.", 'type': 
'invalid_request_error', 'param': 'temperature', 'code': 'unsupported_value'}

I think this should fail gracefully as it's a common attribute sent to other models (makes switching models more difficult if you have to also remember and configure which parameters to send / not send)

References

https://community.openai.com/t/temperature-in-gpt-5-models/1337133/13 - breaking change according to OpenAI community comments

[Kludex]

What does "fail gracefully" means? You want to retry when it has this failure, or not show anything?

[DouweM]

We already disable sampling settings (temperature, presence_penalty, frequency_panelty, logit_bias, logprobs, top_logprobs) for OpenAI reasoning models in openai_model_profile:

pydantic-ai/pydantic_ai_slim/pydantic_ai/profiles/openai.py

Lines 33 to 44 in 5745402

	def openai_model_profile(model_name: str) -> ModelProfile:
	"""Get the model profile for an OpenAI model."""
	is_reasoning_model = model_name.startswith('o')
	# Structured Outputs (output mode 'native') is only supported with the gpt-4o-mini, gpt-4o-mini-2024-07-18, and gpt-4o-2024-08-06 model snapshots and later.
	# We leave it in here for all models because the `default_structured_output_mode` is `'tool'`, so `native` is only used
	# when the user specifically uses the `NativeOutput` marker, so an error from the API is acceptable.
	return OpenAIModelProfile(
	json_schema_transformer=OpenAIJsonSchemaTransformer,
	supports_json_schema_output=True,
	supports_json_object_output=True,
	openai_supports_sampling_settings=not is_reasoning_model,
	)

Looks like we should add gpt-5 as well.

@claude Implement that.

[claude]

Claude finished @DouweM's task —— View job • claude/issue-2482-20250808-1310 • Create PR ➔

---

Implementing GPT-5 temperature parameter fix

Successfully added GPT-5 models to the existing logic that disables sampling settings for OpenAI reasoning models.

Todo List:

• Read existing OpenAI profile code to understand current implementation
• Add gpt-5 models to the logic that disables sampling settings
• Run linting and type checking
• Run tests to ensure changes work
• Commit and push changes
• Provide PR creation URL

Changes Made:

• Updated openai_model_profile() function in pydantic_ai_slim/pydantic_ai/profiles/openai.py:35
• Extended reasoning model detection from model_name.startswith('o') to also include model_name.startswith('gpt-5')
• GPT-5 models now have openai_supports_sampling_settings=False, preventing temperature and other sampling parameters from being sent to the API
• Verified the fix works correctly:
  • gpt-5-mini: openai_supports_sampling_settings=False ✓
  • o1-mini: openai_supports_sampling_settings=False ✓ (unchanged)
  • gpt-4o: openai_supports_sampling_settings=True ✓ (unchanged)

All linting, type checking, and core tests passed. The change gracefully handles the OpenAI API breaking change for GPT-5 models while maintaining backward compatibility.

[gusfraser]

What does "fail gracefully" means? You want to retry when it has this failure, or not show anything?

I accept it wasn't worded as clearly as it could have been - should have been something more like:

"Do not fail when using new OpenAI gpt-5 models just because a temperature parameter has been included. Exclude the temperature parameter from being passed on to OpenAI for gpt-5 models".

[fdb02983rhy]

We already disable sampling settings (temperature, presence_penalty, frequency_panelty, logit_bias, logprobs, top_logprobs) for OpenAI reasoning models in openai_model_profile:

pydantic-ai/pydantic_ai_slim/pydantic_ai/profiles/openai.py

Lines 33 to 44 in 5745402

def openai_model_profile(model_name: str) -> ModelProfile:
"""Get the model profile for an OpenAI model."""
is_reasoning_model = model_name.startswith('o')
# Structured Outputs (output mode 'native') is only supported with the gpt-4o-mini, gpt-4o-mini-2024-07-18, and gpt-4o-2024-08-06 model snapshots and later.
# We leave it in here for all models because the default_structured_output_mode is 'tool', so native is only used
# when the user specifically uses the NativeOutput marker, so an error from the API is acceptable.
return OpenAIModelProfile(
json_schema_transformer=OpenAIJsonSchemaTransformer,
supports_json_schema_output=True,
supports_json_object_output=True,
openai_supports_sampling_settings=not is_reasoning_model,
)
Looks like we should add gpt-5 as well.

@claude Implement that.

Hi, maybe the gpt 5 chat model should be excluded as a special case.