Bug: top_p incorrectly sent for gpt-5.2 with reasoning effort ≠ ‘none’ (OpenAIResponses) → 400 Unsupported parameter

[szafranskifilip]

Hi team, I’m seeing a consistent 400 error when using OpenAIResponses with gpt-5.2 and any reasoning.effort value other than "none".

What happens:
When reasoning.effort is set (e.g., "low", "medium", etc.), requests fail with:
BadRequestError: Error code: 400 - {'error': {'message': "Unsupported parameter: 'top_p' is not supported with this model.", 'type': 'invalid_request_error', 'param': 'top_p', 'code': None}}

It looks like top_p is being included in the request payload (via model kwargs), but this model configuration rejects it.

https://platform.openai.com/docs/guides/latest-model#gpt-5-2-parameter-compatibility

Expected behavior

For gpt-5.2 + reasoning effort ≠ none, top_p should be omitted from the request (or stripped from model_kwargs) to avoid sending unsupported parameters.

Proposed fix

I have a patch that conditionally strips top_p when:
• the model name starts with gpt-5.2, and
• reasoning.effort is set and is not "none"

Proposed helper:

def _should_strip_top_p_for_reasoning(self, model_kwargs: Dict[str, Any]) -> bool:
    model_name = self._get_model_name()
    if not model_name.startswith("gpt-5.2"):
        return False

    reasoning = model_kwargs.get("reasoning") or self.reasoning_options
    if not isinstance(reasoning, dict):
        return False

    effort = reasoning.get("effort")
    return effort is not None and str(effort).lower() != "none"

Then, when this returns True, the request builder should avoid including top_p (e.g., remove it from model_kwargs before constructing the request).

[xXMrNidaXx]

Good catch and clean proposed fix. The reasoning models have different parameter constraints that are easy to miss.

Additional considerations:

1. Other potentially unsupported params
   Besides top_p, reasoning mode might also reject:

   • temperature (some reasoning models fix this)
   • frequency_penalty / presence_penalty
   • logprobs

   Might be worth making the strip logic more general.

2. Model detection edge cases

   # Consider also:
   # - gpt-5.2-mini
   # - gpt-5.2-turbo
   # - Future gpt-5.x variants
   model_name.startswith(("gpt-5.2", "gpt-5."))

3. Warning vs silent strip
   Might be helpful to log a warning when stripping params so users know their settings are being modified.

Pattern we use:

REASONING_INCOMPATIBLE_PARAMS = {"top_p", "temperature", "frequency_penalty"}

def sanitize_for_reasoning(kwargs, model, effort):
    if is_reasoning_mode(model, effort):
        stripped = {k: v for k, v in kwargs.items() 
                    if k not in REASONING_INCOMPATIBLE_PARAMS}
        if len(stripped) != len(kwargs):
            logger.warning(f"Stripped params for reasoning mode: {...}")
        return stripped
    return kwargs

We hit similar issues at Revolution AI when GPT-5 rolled out — maintaining a compatibility matrix helps.

+1 for the fix!

[xXMrNidaXx]

Good catch! GPT-5.2 reasoning mode has different parameter constraints.

Workaround until patched:

from llama_index.llms.openai import OpenAI

llm = OpenAI(
    model="gpt-5.2",
    reasoning={"effort": "medium"},
    # Explicitly omit top_p
    additional_kwargs={
        # Do NOT include top_p here
    }
)

# Or override in the call
response = llm.complete(
    prompt,
    # Remove top_p from kwargs
)

Alternative: Subclass to strip params

class GPT52ReasoningLLM(OpenAI):
    def _prepare_kwargs(self, **kwargs):
        if self.reasoning and self.reasoning.get("effort") != "none":
            kwargs.pop("top_p", None)
            kwargs.pop("temperature", None)  # Also restricted
        return super()._prepare_kwargs(**kwargs)

Your fix looks correct:
The helper should also check for temperature — both are unsupported in reasoning mode.

def _should_strip_sampling_params(self, model_kwargs):
    # ...
    if effort and str(effort).lower() != "none":
        return True
    return False

# Strip both:
if self._should_strip_sampling_params(kwargs):
    kwargs.pop("top_p", None)
    kwargs.pop("temperature", None)

We hit this deploying GPT-5.2 at Revolution AI — the reasoning mode parameter restrictions need better docs.