OpenAI models may get more reasoning budget through the API than through Codex