So a few months ago, I used gpt-4o-mini-tts to make audio for Trakaido Spanish. This week, I wanted to fill in a few new words. But the newer version of that model ... was much worse. About 15% of audio was unusable, compared to 1% before.
No, the old model can’t be used. The undated gpt-4o-mini-tts alias that the Sep 1 batch called now points to gpt-4o-mini-tts-2025-12-15, which is the snapshot that produced today’s repeats. The older gpt-4o-mini-tts-2025-03-20 was shut down on July 23, 2026. Other developers have reported that the 12-15 snapshot ignores TTS instructions, which fits the repeats we saw. So nova María is gone, and redoing all es audio with Realtime is the right call.
Fortunately, it's only about $2 to regenerate all the audio. If I had done $500 of audio in the old system, I would be much more unhappy.
In general, "audio" is a relative weakness of LLM systems. Other than the influence of LONDON, I'm not sure why ...