GPT-Transcribe Replaces Whisper — 52% Fewer Errors, 25% Less
- What happened
- OpenAI released GPT-Transcribe (batch) and GPT-Live-Transcribe (streaming) to the API, replacing Whisper with 52% fewer errors at $4.50/1K minutes — a 25% price cut from GPT-4o Transcribe's $6.00.
- Why it matters
- Context-aware prompting, keyword hints, and language hints fix Whisper's three biggest failure modes — specialized vocabulary, accented speech, and noisy audio — while costing less than the models they replace.
- What to do
- If you're on Whisper or GPT-4o Transcribe, migrate — the API shape is the same and accuracy is dramatically better. If you need real-time streaming, GPT-Live-Transcribe is the new default. Stay on gpt-4o-transcribe-diarize if you need speaker labels or timestamps.
OpenAI just made Whisper obsolete. On July 28, the company shipped GPT-Transcribe and GPT-Live-Transcribe to the API — two speech-to-text models that cut transcription errors by 52% while costing 25% less than the models they replace.
GPT-Transcribe handles batch and async file transcription at $0.0045 per minute ($4.50 per 1,000 minutes), down from GPT-4o Transcribe's $0.006. GPT-Live-Transcribe streams real-time transcript deltas at $0.017 per minute with tunable latency — built for live captioning, meeting assistants, and voice agents where sub-second response matters more than per-minute cost.
What happened
OpenAI released both models through the existing /v1/audio/transcriptions and Realtime API endpoints. GPT-Transcribe targets completed audio files; GPT-Live-Transcribe connects via WebSocket or WebRTC for continuous streaming. Neither model replaces ChatGPT's voice features — this is an API-only launch.
The headline number is a 52% error reduction on Common Voice across 22 languages: GPT-Transcribe scored 19.27% transcription error rate versus whisper-1's 40.37% (AlphaSignal(opens in new tab), 2026). On independent benchmarks, GPT-Transcribe scores 3.31% AA-WER (Artificial Analysis(opens in new tab), 2026) — a 0.7 percentage-point improvement over GPT-4o Transcribe.
What makes these models different from Whisper is context awareness. Both accept three kinds of input that traditional ASR ignores:
| Input type | What it does | Example |
|---|---|---|
| Text prompt | Describes the recording's topic and expected vocabulary | "Medical consultation about cardiac symptoms" |
| Keyword list | Boosts recognition of domain-specific terms | ["myocardial", "troponin", "echocardiogram"] |
| Language hints | Improves accuracy on accented or mixed-language speech | Spanish-dominant speaker in an English context |
These features directly address where Whisper most consistently failed: specialized terminology (medical, legal, technical), accented speech, and noisy environments. OpenAI's Context Aware ASR benchmark shows context injection boosts semantic accuracy by 3–6 percentage points.
Why it matters
A 52% error reduction is a generational leap for speech-to-text. For anyone building on Whisper — meeting summarizers, call center analytics, content indexing, accessibility tools — this is a drop-in upgrade that costs less and performs dramatically better. The same API shape and endpoint means migration is a model-name swap.
The competitive picture is straightforward: these aren't LLMs and don't compete on reasoning or coding. But for the specific job of turning audio into text, OpenAI just set a new baseline. The recommended verdicts on both models reflect the combination of accuracy gains, lower pricing, and context-aware prompting.
The gaps are real, though. Neither model supports speaker diarization, word-level timestamps, SRT/VTT output, or English translation. Teams that need speaker labels should stay on gpt-4o-transcribe-diarize; those needing timestamps still use whisper-1.
What changes for you
If you're on Whisper or GPT-4o Transcribe, migrate. GPT-Transcribe accepts the same API shape at lower cost with higher accuracy. The model name changes; your request structure doesn't.
If you need real-time streaming, GPT-Live-Transcribe is the new default. It's roughly 3.8x the per-minute cost of batch GPT-Transcribe, but delivers sub-second transcript deltas that live captioning and voice agents require.
Watch the gaps. No diarization, no timestamps, no SRT/VTT. These features haven't disappeared — they're just split into separate model paths (gpt-4o-transcribe-diarize, whisper-1). Plan your migration around what your pipeline actually needs.
FAQ
Is Whisper dead? For most new builds, yes. OpenAI is positioning GPT-Transcribe as the recommended starting point, and the 52% error reduction on Common Voice makes Whisper hard to justify for greenfield work. However, whisper-1 remains available for workloads that depend on word timestamps or English translation, which the new models don't support.
Do these models work in ChatGPT? No. The launch is API-only. ChatGPT's voice features run on separate models (GPT-Live-1 and GPT-Live-1 mini). These transcription models are for developers building their own voice pipelines.
How does context-aware prompting actually help? Instead of treating every audio file as a blank slate, you provide hints about what's being discussed. A medical transcription with keyword hints like "troponin" and "myocardial" will recognize those terms far more reliably than a generic model. The improvement is most dramatic on domain-specific audio where Whisper would routinely mangle specialized vocabulary.
What to do
- 1 Add context prompts and keyword lists to your requests — the accuracy gain on domain-specific audio is the biggest differentiator from Whisper
- 2 Swap whisper-1 to gpt-transcribe in your transcription endpoint — same API shape, 52% fewer errors, 25% lower cost
- 3 Replace gpt-realtime-whisper with gpt-live-transcribe for live captioning and voice agent pipelines
- 4 Keep gpt-4o-transcribe-diarize if your pipeline depends on speaker labels or word-level timestamps
Affected tools & models
Never need to catch up again
The weekly delta — only verdict changes and act-now items. No digest filler.