OpenAI's New Transcriber Improves, Still Trails Three Rivals

OpenAI shipped two speech-to-text models through its API this week: GPT Transcribe for pre-recorded audio and GPT Live Transcribe for streaming. The batch model, GPT Transcribe, processes audio at roughly 34 times its own playback speed. The second model handles live audio streams and is tuned to keep latency low, a separate engineering problem from transcribing a finished file.

Artificial Analysis, the firm behind the AA-WER speech recognition benchmark, measured GPT Transcribe’s word error rate at 3.31 percent. The error rate marks a 0.7-point drop from GPT-4o Transcribe, the model it replaces from a year earlier. Price fell too: a minute of audio now costs $0.0045, a 25 percent cut from the prior rate. Both new models accept text as transcription context, keyword lists, and multiple input languages.

The AA-WER leaderboard still ranks three vendors ahead of OpenAI on raw accuracy. ElevenLabs’ Scribe v2 sits at 2.3 percent, Google’s Gemini 3 Pro follows at 2.9 percent, and Mistral’s Voxtral Small comes in at 3 percent, each beating GPT Transcribe’s 3.31 percent. Mistral has also pushed on price, listing its newer Voxtral Transcribe V2 at $0.003 per minute of audio, undercutting OpenAI’s new rate by a third.

Those numbers come from Artificial Analysis’s own benchmark run, not from any of the vendors involved, which gives the ranking more weight than a company’s self-reported figures. Word error rate is still one input into a purchasing decision, not the whole case. Latency matters for voice agents and live captioning, which is why OpenAI built a separate streaming model instead of adapting the batch one. Diarization (labeling which speaker said what), the breadth of supported languages, and per-minute pricing at high volume all factor into total cost.

A call center transcribing millions of minutes a month might take Mistral’s slightly higher error rate in exchange for the lowest price on the board. A legal or medical transcription vendor might pay ElevenLabs’ premium because a single missed word carries more risk there than the price difference does. GPT Transcribe’s gains narrow that calculus for OpenAI’s existing customers without closing it.

The two models extend OpenAI’s Realtime lineup, which already includes the streaming transcription model GPT-Realtime-Whisper, and OpenAI published integration details in its Transcription Guide for developers evaluating a switch. The Decoder, reporting on July 29, credited Artificial Analysis’s AA-WER rankings for the comparison figures.

Teams currently running GPT-4o Transcribe in production should benchmark their own audio against Scribe v2 and Gemini 3 Pro before renewing, since a lower word error rate can still lose to a better price or better diarization depending on the workload.

The Decoder (Jonathan Kemper) reported this on July 29, 2026, citing word-error-rate figures from Artificial Analysis’s AA-WER benchmark.