OpenAI Ships GPT-Transcribe at $0.0045/Min and GPT-Live-Transcribe at $0.017/Min
OpenAI has shipped two new transcription models in its API: GPT-Transcribe, built for asynchronous file and batch processing, and GPT-Live-Transcribe, built for streaming low-latency transcription of live audio.
Both replace the prior generation gpt-realtime-whisper models and represent the first transcription-specific models in OpenAI’s GPT-branded lineup — distinct from the multimodal audio tiers of GPT-Realtime.
Models and Pricing
| Model | Use Case | Price |
|---|---|---|
| GPT-Transcribe | Async file, batch, Realtime committed turns | $0.0045/min |
| GPT-Live-Transcribe | Live streaming, low-latency deltas | $0.017/min |
| gpt-4o-transcribe (prior gen) | Async file | $0.006/min |
| gpt-realtime-whisper (prior gen) | Streaming | $0.017/min |
GPT-Transcribe at $0.0045/min is 25% cheaper than the outgoing gpt-4o-transcribe ($0.006/min). GPT-Live-Transcribe matches the prior gpt-realtime-whisper price while posting lower error rates.
Benchmark Numbers
On the Real-World Audio Recording benchmark across nine languages, GPT-Live-Transcribe achieved a 9.60% transcription error rate, compared to 11.65% for GPT-Realtime-Whisper-1 — a 17.6% improvement. Across 22 languages on Common Voice, the margin is narrower: 19.70% vs 20.33%.
On OpenAI’s new Context Aware Automatic Speech Recognition benchmark, GPT-Live-Transcribe’s semantic accuracy rose from 38.5% without free-form context to 44.6% with it — a 16-percentage-point lift from a single paragraph of background text.
The models target audio failure modes that prior generations handled poorly: short phrases, numeric strings, domain-specific terminology, multilingual speaker switches, and high-background-noise environments.
Context API
Both models accept structured context at inference time:
- Free-form context: unstructured background about the recording (meeting topic, speaker roles, industry)
- Keyword hints: names, acronyms, product terms that generic transcription models regularize to common words
- Language hints: expected input languages for multilingual audio
- Earlier turns: prior transcript segments as context (GPT-Live-Transcribe)
The context mechanism is the sharpest practical differentiation from Whisper-based models. Enterprise transcription workloads — earnings calls, medical consultations, legal depositions — routinely fail on proper nouns and domain terminology. The keyword hint API addresses this directly without fine-tuning.
Positioning
The two-model split reflects a deliberate product decision: async and streaming transcription have different latency budgets, infrastructure requirements, and pricing sensitivities. Separating them lets OpenAI optimize each independently and lets developers avoid paying streaming inference rates for batch jobs.
GPT-Transcribe also connects to the Realtime API: it handles committed turns in Realtime WebSocket sessions, making it viable as a drop-in component inside voice agent pipelines that need higher-accuracy transcription than the base Realtime model provides.
Whisper, the open-weight baseline OpenAI released in 2022, remains available for self-hosted inference. The new API models target production deployments where latency guarantees and context-aware accuracy matter more than hosting cost.