New OpenAI speech models lag behind rivals in accuracy, cost

OpenAI launched GPT Transcribe and GPT Live Transcribe, speech recognition models that transcribe audio into text. GPT Transcribe transcribes audio 34x faster than realtime, achieves a WER of 3.31% (improved by 0.7 points compared to the previous version), and is 25% cheaper at $0.0045 / minute. GPT Live Transcribe transcribes audio in realtime and streams audio.

GPT Transcribe and GPT Live Transcribe can take contextual text as input for improved accuracy, allow specifying keywords, and support multiple languages. On the AA-WER benchmark, ElevenLabs Scribe v2 achieves a WER of 2.3%, Google Gemini 3 Pro achieves a WER of 2.9%, Mistral Voxtral Small achieves a WER of 3.0% and is cheaper ($0.003 / min). OpenAI lags behind all three competitors in terms of both accuracy and cost.

In contrast to chatbots, where OpenAI is the industry leader, the field of speech recognition is dominated by other players. We see that no lab can be the best in everything, and the AI landscape becomes increasingly fragmented, providing opportunities for niche players like ElevenLabs or Mistral to compete effectively with OpenAI.