OpenAI reasoning models reject max_tokens and require max_completion_tokens instead

The o-series and later reasoning models do not accept the max_tokens parameter that Chat Completions has used since the beginning. Requests carrying it are rejected with an unsupported-parameter error naming max_completion_tokens as the replacement. The rename exists because the new parameter bounds reasoning tokens plus visible output, whereas max_tokens bounded only visible output.

Several other long-standing parameters are also unsupported or fixed on these models, including temperature and top_p, so a generic client that always sends a full parameter set will fail on model swap even though the prompt is fine.

The practical fix is to build the request body per model family rather than sending one superset and hoping unknown fields are ignored, because they are not. When adding a new model to a router, send one trivial request with your standard parameter set and read the error before wiring it into traffic.