OpenAI’s new transcription models improve but trail rivals on errors

OpenAI has released GPT Transcribe and GPT Live Transcribe, two speech-recognition models available through its API.

The Decoder reports that the new models improve on OpenAI’s previous transcription offering, but do not beat ElevenLabs, Google, or Mistral on error rates. That makes the release useful for OpenAI developers while leaving the broader speech-to-text race unsettled.

Transcription accuracy matters because small errors can change names, instructions, medical details, or legal meaning. Live transcription adds another constraint: systems must return text quickly enough for real-time use while still handling accents, background noise, and specialized vocabulary.

For builders already using OpenAI’s API, the new models may simplify product architecture by keeping speech and language workflows under one provider. But teams choosing a transcription engine on accuracy alone still have reason to compare alternatives before switching.