TTS with a question... the audio back has the answer

Thanks for the clear repro. This does look like a behavior issue with gpt-4o-mini-tts-2025-12-15, where some prompt-like input can get interpreted instead of spoken verbatim. The quickest workaround is to pin the Speech API model to gpt-4o-mini-tts-2025-03-20, which behaved correctly in testing for this case. If you’re open to a larger change, gpt-audio-1.5 is also worth testing, and realtime models can work for voice use cases too, but those are not a drop-in replacement for the current /v1/audio/speech request shape.