There is a workaround to this. Just forbid it to route the request to anything other than GPT-5. But again the lack of transparency of the model used, creates confusion and leads to conflicting opinions on gpt-5 models performance.