OpenAI o-series / GPT-5: `Unsupported parameter: 'max_tokens'` (use `max_completion_tokens`) and `temperature` only supports 1

发生场景

Source(s): https://github.com/zed-industries/zed/issues/33103 , https://github.com/BerriAI/litellm/issues/13381 Signal: GitHub search finds about 875 issues with this exact error (tools broke when o1/o3/o4-mini and later GPT-5 shipped). Status: reported working by source, not reproduced by us.

问题

Calls that worked with GPT-4-class models fail with 400 on reasoning models.

报错原文（已脱敏/规范化）

Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead. ; Unsupported value: 'temperature' does not support 0.2 with this model. Only the default (1) value is supported.

不是这个问题时 / 边界

The model is not a reasoning model, or the error names a different parameter. (our inference: gpt-5-codex style models that only exist on the Responses API fail with "not accessible via the /chat/completions endpoint" instead; that needs the Responses API, not this change.)

根因

Reasoning models reject max_tokens (they take max_completion_tokens) and only accept the default temperature.

修复步骤

1. In your own code: send max_completion_tokens instead of max_tokens, and leave out temperature (or set it to 1).
2. In tools: update. A Zed user (not the reporter) said the preview build that maps the parameter "works just fine". LiteLLM fixed the mapping for GPT-5 (PR #13390); its drop_params option drops unsupported values.

如何验证

The same request returns 200.