What problem does this feature solve?

With the release of the gpt-5 series of models, the cost can go way down for some use cases (mine in particular - web testing with playwright using the gpt-4.1 models) but the API has changed and now "max_tokens" has been replaced by "max_completion_tokens" and throws an "Unsupported parameter" error when used.

What does the proposed API look like?

Perhaps we add a "GPT-5" mode that changes that call in the common config to use that parameter name when this mode is active?