OpenAI ships GPT-Transcribe and GPT-Live-Transcribe APIs
Two new speech-to-text models replace Whisper as OpenAI's recommended default, with steerable context inputs and pricing from $0.0045 per minute.
OpenAI released two speech-to-text models in its API on 28 July, positioning them as the recommended path for new work over both Whisper and the gpt-4o-transcribe family.
The split
GPT-Live-Transcribe targets low-latency streaming and is exposed only through the realtime transcription endpoint, priced at $0.017 per minute of audio with tunable latency. GPT-Transcribe handles completed files and batch workloads at $0.0045 per minute. OpenAI says both improve on real-world audio — accents, code-switching, numbers, short utterances, domain jargon and heavy background noise — and reports error rates on the async model at less than half Whisper's on its internal comparisons.
Steerable context
The more consequential change is that transcription becomes promptable. Developers can pass free-form context describing the recording, a keyword list for proper nouns and domain terms, expected input languages, and formatting expectations. OpenAI's published figure on the Context Aware ASR benchmark is that semantic accuracy rises from 41.6% with no free-form context to 45.2% with it — a modest absolute gain, but on the exact failure mode that has made ASR expensive to productise: a model that transcribes a clinical consultation or a sales call without knowing the drug names, ticker symbols or internal product codes involved. Until now that was patched with post-hoc LLM correction passes. OpenAI has published a migration cookbook for teams moving off Whisper.
Why it matters
Whisper's 2022 open release effectively set the floor for the entire speech-to-text market, and OpenAI has now formally deprioritised it in favour of closed API models with per-minute billing. For the voice-agent stack — where transcription sits upstream of every downstream decision and a misheard name propagates through the whole turn — steerable context removes a layer of glue code that most production deployments currently maintain themselves. It also lands the same week Fish Audio raised $52M for creator- and enterprise-facing voice models, a reminder that the audio layer is now contested by well-funded specialists rather than treated as solved infrastructure. The pricing gap between the streaming and batch tiers is nearly 4x, which will push architects to keep as much transcription off the realtime path as latency budgets allow.