OpenAI’s new speech-to-text model, GPT Transcribe, turned up on OpenRouter priced at less than half a cent a minute. That is about 27 cents to transcribe an hour of audio. There is just one odd detail.
On the 5th of August, GPT Transcribe appeared on OpenRouter, the marketplace that resells access to models from dozens of labs, at $0.0045 per minute of audio. Do the sums and an hour of recording costs roughly 27 cents to turn into text. For anyone who records interviews, podcasts, voice notes or client calls, that is cheap enough to stop thinking about.
Speech-to-text (STT, the tech that turns spoken audio into written words) has become one of the most useful and least glamorous corners of the AI stack. GPT Transcribe is built for three jobs: transcribing recorded audio files, streaming a file as it transcribes, and handling ‘committed’ turns in a live Realtime conversation. It also takes free-form context, keyword hints and multiple language hints, which in practice means you can feed it a list of odd names, product terms or acronyms and it will stop mangling them.
The odd detail
Here is what the excitable write-ups skip. As of the listing, there was no first-party OpenAI launch post and no OpenAI price page to check that 27-cents figure against. The rate lives on OpenRouter, not on OpenAI’s own pricing surface. That matters, because a marketplace rate belongs to the marketplace. It can sit above or below whatever the vendor eventually publishes, and it can move without warning. Treat $0.0045 a minute as a real price you can pay today, not as OpenAI’s confirmed list rate.
For scale, OpenAI’s older GPT-4o Transcribe is priced by tokens ($2.50 in, $10 out per million), which makes per-hour costs fiddly to predict. A flat per-minute rate is far easier to budget, and that alone makes GPT Transcribe worth a look for anyone running transcription at any real volume.
The angle: the little guy wins here, genuinely. Twenty-seven cents an hour with keyword hints is professional-grade transcription for the price of nothing. Just do not build a business plan on a marketplace banner. Test it on your own messy audio (accents, crosstalk, background noise) before you retire whatever you use now, and keep an eye out for OpenAI’s own price page landing later and quoting a different number.
Did you know: OpenAI’s older transcription model, Whisper, was released free and open-source in 2022, and it is still the workhorse behind a huge share of the transcription apps you have used without ever knowing its name.
Related on Top Tool Stack: The price of thinking fell 80% · Grok’s 8-cent voice model