Step 1
Identify the target model and reasoning mode
Do not treat temperature support as universal. Check the exact model, endpoint, and reasoning setting used by the failing request.
OpenAI parameter error
“temperature is not supported”: check reasoning settings before retrying
On current OpenAI reasoning models, sampling controls can be restricted depending on reasoning effort. OpenAI’s migration guidance says to remove temperature, top_p, and top_logprobs when reasoning effort is not none.
What to check
Step 2
Remove unsupported sampling fields
When the target model is running with reasoning enabled and the docs mark sampling controls unsupported, remove those fields from the request builder rather than forcing an old configuration onto the new model.
Step 3
Regression-test prompts and tools
Removing sampling parameters can alter behavior. Run representative prompts, structured outputs, and tool calls so the migration is measured by task success instead of by whether the request merely stops erroring.
The exact supported parameter set is model-specific. Verify the current model documentation before publishing a broad compatibility change.
Check the rest of the repository
Fixing the first error does not prove the rest of the codebase is clear. Sunset scans source text for scheduled OpenAI shutdown risk and shows severity before you pay.