Every generation request to our newest, most capable OpenAI tier was failing. The real error: “Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.”
The parameter itself had been standard across every OpenAI model we'd used up to this point. The newest model generation rejects it outright rather than accepting it with a deprecation warning — a hard break, not a soft one.
The fix applied everywhere we called this provider, not just the one failing path — the new parameter name works universally, including on older models, so there was no reason to branch by model.