OpenAI API Update
GPT-Transcribe splits live vs. recorded
On July 29, 2026, OpenAI overhauled its API transcription lineup, replacing the do-everything whisper-1 with a two-model split: "GPT-Transcribe" for recorded audio and "GPT-Live-Transcribe" for live streams. Splitting by workload is, in itself, the first hint of a real design philosophy behind a month of otherwise scattershot voice releases.
Scattered releases finally line up
OpenAI has shipped voice-related news almost weekly this July. Seen in sequence, this split reads differently than a standalone feature drop.
July 6 brought GPT-Realtime-2.1 / 2.1-mini (reasoning added to the Realtime API, p95 latency cut by 25%+); July 8 brought GPT-Live, the full-duplex voice model behind ChatGPT's voice mode; July 22 brought Presence, an enterprise voice-agent deployment service claiming 75% of inbound calls on OpenAI's own support line get resolved without a human. Taken individually, these look like a feature-of-the-week list. The July 29 transcription overhaul is what finally draws a line through them.
The key move is that OpenAI explicitly split two distinct voice workloads — "live" and "recorded" — into separate models rather than forcing both through one. GPT-Live carries ChatGPT's conversational experience, GPT-Realtime-2.1 carries low-latency API dialogue, and now these two models carry transcription. A product line built around matching the model to the job is finally visible. Read this less as "another weekly feature" and more as evidence that the voice stack is being rebuilt as a coherent architecture.
More than 3x apart on price, better on error rate
The price gap between the two models maps directly onto the use-case gap. Accuracy is also meaningfully improved over whisper-1.
At a rough 150 yen-per-dollar conversion, GPT-Transcribe works out to roughly ¥0.68/min and GPT-Live-Transcribe to roughly ¥2.55/min (both approximate — actual cost depends on the exchange rate at billing time). The live model costs more than 3.7x the recorded one, and that gap is essentially the price of always-on, low-latency processing showing up directly in the rate card.
On accuracy, OpenAI's official model documentation states both models accept "context" inputs — free-form prompts, keyword hints, and language hints — to improve handling of accents, numbers, specialized terminology, and noisy audio. Independent reporting cites GPT-Transcribe's word-error-rate (WER) on the Common Voice dataset across 22 languages dropping from roughly 40% under the older whisper-1 to roughly 19% — a clear improvement over the prior model, as reported.
From "one model for everything" to "one model per job"
| whisper-1 (old, single model) | GPT-Transcribe / Live-Transcribe (new, split) |
|---|---|
| Live and recorded handled by the same model | Separate low-latency live and async recorded models |
| No per-use-case pricing | Clear split: $0.0045/min vs. $0.017/min |
| No context input (prompts/keywords) | Supports prompt, keyword, and language hints |
| Common Voice WER ~40% (reported) | GPT-Transcribe WER improved to ~19% (reported) |
Who this helps, and how
Engineers: pick the right model
Latency-critical uses like meeting captions or live-stream subtitles want GPT-Live-Transcribe; batch jobs like podcast or call-center transcription want GPT-Transcribe. Mixing the two on the same endpoint costs you on both price and latency.
Business: rethink the cost structure
Splitting price by workload lets teams that were paying one flat whisper-1 rate separate their "live" spend from their "batch" spend. If you don't use live transcription, sticking to GPT-Transcribe alone simply lowers cost.
PMs: deciding on the whisper-1 migration
whisper-1 stays available for now, but the new models lead on both accuracy and features. The migration effort is less about code changes and more about re-evaluating accuracy against your own audio first.
This isn't a story about being the world's most accurate.
It's a story about OpenAI closing its own gap.
The counterargument, and the limits
Independent outlet The Decoder adds an important caveat here. GPT-Transcribe is an improvement over OpenAI's own prior model, but on raw error-rate benchmarks it still trails ElevenLabs, Google, and Mistral. This isn't a "best transcription in the industry" announcement — it's OpenAI catching up to a gap in its own lineup.
For use cases where transcription accuracy is directly load-bearing for the business — medical dictation, financial call review — that gap may not be negligible. Whether the convenience of staying inside OpenAI's ecosystem outweighs the raw-accuracy edge held by other vendors depends on the workload.
What to do next
Benchmark against your own audio
If you're on whisper-1 today, measure GPT-Transcribe's accuracy against your own real audio before migrating — the reported WER improvement will vary by language and recording conditions.
Split live and batch traffic by model
Route low-latency uses like captions and subtitles to GPT-Live-Transcribe, and batch workloads to GPT-Transcribe, rather than defaulting everything to one model.
Compare ElevenLabs and others if raw accuracy matters most
If error rate is the top priority, weigh ElevenLabs, Google, and Mistral on price and accuracy too. Ecosystem consistency inside OpenAI and best-in-class raw accuracy are a real trade-off here.