GPT Transcribe: OpenAI's Accuracy Gains Still Trail ElevenLabs, Google, and Mistral

OpenAI’s GPT Transcribe is a real upgrade over the model it replaces: a 0.7 percentage point cut in word-error rate and a 25 percent price drop to $0.0045 per minute of audio. What it did not do is take the accuracy crown. On independent benchmarking, GPT Transcribe still trails three rivals it was supposed to be closing in on.

Why This Comparison Matters

Word-error rate (WER) is the metric that decides whether a transcription API survives contact with production traffic. A model that mishears one word in twenty is tolerable for a rough draft; it is not tolerable for medical dictation, legal transcripts, or a voice agent that has to act correctly on what it just heard. That’s why the AA-WER benchmark run by Artificial Analysis, which The Decoder covered in detail, has become a standard reference point for developers choosing a speech-to-text provider — much the same way real-time voice models get judged on response latency, not transcript quality alone.

What Makes It Different

GPT Transcribe’s headline number is progress, not victory. According to the AA-WER results, GPT Transcribe posts a 3.31 percent error rate, down from the 4.01 percent of its year-old predecessor, GPT-4o Transcribe. That’s a meaningful jump in one release cycle. But it still lands behind every other major provider on the same leaderboard: ElevenLabs Scribe v2 leads at 2.3 percent, Google’s Gemini 3 Pro follows at 2.9 percent, and Mistral’s Voxtral Small comes in at 3 percent.

Pricing complicates the picture further. GPT Transcribe’s 25 percent cut brings it to $0.0045 per minute, but Mistral has already undercut that with Voxtral Transcribe V2 starting at $0.003 per minute — a model that also posts a lower error rate. OpenAI’s move narrows the accuracy gap without closing the price gap.

The Catch / Who It’s For

The catch is straightforward: on the metric OpenAI itself is being measured against, it’s still fourth out of four. The gap to ElevenLabs is roughly a full percentage point of WER, which is significant at this end of the leaderboard, where every provider is fighting over fractions of a percent. Teams building voice agents, call-center transcription, or captioning pipelines where accuracy is the deciding factor have little reason to switch to GPT Transcribe on merit alone.

Where it does make sense is for teams already standing on OpenAI’s infrastructure — using GPT-4o or the Realtime API elsewhere in their stack — who value one vendor relationship and one bill over squeezing out the last percentage point of accuracy. For them, a cheaper, more accurate-than-before option inside a stack they already run is a legitimate reason to stay put rather than integrate a second provider.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.