OpenAI Whisper API pricing 2026, model by model

OpenAI’s Whisper API costs $0.006 per minute of audio, or $0.36 an hour. That price hasn’t changed since the API launched in March 2023, which is why so many transcription bills were built on it.

Two things changed in 2026. OpenAI launched cheaper, more accurate transcription models, and on August 26 it deprecated whisper-1 along with every GPT-4o transcription model. They leave the API on February 26, 2027. So the useful question is no longer only “what does Whisper cost?” It’s “what will this audio cost after February, and on which model?”

Below are OpenAI’s current transcription prices, read from OpenAI’s pricing page on October 11, 2026, then what the shutdown breaks, what real workloads cost, and the alternatives. We build Cartesia, which sells speech to text, so we have an interest here. Every OpenAI number is OpenAI’s own, and our prices are at the end so you can check both.

OpenAI transcription pricing

The ”~” rows are OpenAI’s own estimates. Those three models bill by token ($2.50 per million audio input tokens and $10 per million output tokens for gpt-4o-transcribe, half that for mini), so the real per-minute cost depends on how much is said. whisper-1 and the gpt-transcribe family bill by audio duration, which is easier to budget.

A few things the table doesn’t show:

• Files max out at 25 MB. The file transcription guide accepts mp3, mp4, mpeg, mpga, m4a, wav and webm up to 25 MB, so an hour of WAV needs splitting or compressing first. Splitting a recording is how word errors sneak in at the seams.
• Live costs almost four times as much as files. Streaming transcription runs $0.017 a minute against $0.0045 for the same minute uploaded later. If you don’t need words while the person is talking, don’t pay for it.
• Regional processing costs extra. OpenAI adds a 10% uplift on data-residency endpoints for eligible models released on or after March 5, 2026.
• The newer models have no free tier. The model pages for gpt-transcribe and both live models list free-tier accounts as not supported. You pay from the first minute.

The February 2027 shutdown

The deprecation notice is short. Four models go, and two replace them:

For plain transcription the move is mostly a model string and a price cut. gpt-transcribe takes a languages array instead of the single language field, adds keywords for product names and IDs, and can stream text while it works through a file.

The trouble is the jobs Whisper did that nothing replaces yet. OpenAI’s own transcription guide still routes three of them to models on the removal list:

• Word timestamps and srt or vtt subtitles go to whisper-1.
• Translation into English goes to whisper-1 through the translations endpoint.
• Speaker labels go to gpt-4o-transcribe-diarize.

As of October 11, 2026, OpenAI hasn’t named a successor for any of them. If your product makes captions, aligns a transcript to video, or splits a call by speaker, the replacement isn’t in the notice. That’s the part of the migration to start on now, not in February.

What a real workload costs

At list prices, before any data-residency uplift:

Transcribing 10,000 hours of recorded calls a month (600,000 minutes). On whisper-1 that’s $3,600. Moving to gpt-transcribe makes it $2,700. gpt-4o-mini-transcribe would be about $1,800, but it leaves in February too, so that saving lasts a few months at most.

A voice agent listening to 10,000 call minutes a month. Live transcription on gpt-live-transcribe is $170. That sounds small next to the batch bill, and it is: for an agent, transcription is rarely the expensive line. The cost that matters is what a misheard account number does to the rest of the call.

Is Whisper still accurate enough?

Whisper set the bar in 2022. It no longer sits near the top. Artificial Analysis measured these word error rates on recorded audio in October 2026 (lower is better):

Two readings of that table. Inside OpenAI, gpt-transcribe is both cheaper and more accurate than whisper-1, so for plain transcription there’s no reason to wait for February. Outside OpenAI, Groq’s hosted Whisper large-v3 Turbo costs about a ninth of whisper-1 for half a point more error. That’s a fine trade for archives and a poor one for anything a customer reads.

Word error rate is a diagnostic, not a verdict (we wrote a whole guide on that). Run your own hardest recordings through two or three finalists before you migrate.

Running Whisper yourself

The open-source Whisper model is MIT-licensed and isn’t part of the shutdown. If your audio is recorded, you have spare GPUs, and you need the features OpenAI is dropping (timestamps, subtitles, translation into English), self-hosting Whisper large-v3 keeps all three. Hosted Whisper from inference providers is the same model without the GPUs. Either way you own chunking, silence handling and upgrades.

Self-hosting is a poor fit for a live conversation. Whisper was built for 30-second windows of finished audio. A streaming wrapper around it means your own endpointing, your own handling of silence hallucinations, and extra round-trip latency on every turn.

Where Cartesia fits

Cartesia has two speech-to-text models, and they do different jobs.

For voice agents, Ink-2. It streams as the caller speaks and detects the end of a turn itself, so the agent doesn’t need a separate VAD. On the Artificial Analysis streaming benchmark in October 2026, Ink-2 with its built-in turn detection has a 3.4% word error rate, against 3.9% for GPT Live Transcribe and 4.9% for GPT Realtime Whisper. At Scale-plan rates it costs about $0.40 an hour, against $1.02 for either OpenAI live model. Ink-2 covers English, French, Hindi, Japanese and Spanish today. For other languages live, test a multilingual model such as Deepgram Nova-3 alongside OpenAI’s.

For recorded files, ink-whisper. It’s our older Whisper-family model, which we still serve on the batch endpoint: one upload, no file splitting, word timestamps, around 100 languages. Cartesia doesn’t offer translation into English or speaker labels, so it covers one of the three gaps the shutdown leaves, not all three.

Cartesia’s pricing bills both from the same credits as text to speech:

Plans start free with 20,000 credits a month, enough for about 11 hours of batch transcription. Silence is billed as audio. Our plans are prepaid subscriptions rather than pure pay-per-minute, so check the pricing page for the plan that matches your volume.

If you transcribe recorded audio and need timestamps after February, run a few files through ink-whisper. If you’re building a voice agent, try Ink-2 in the playground on your own calls. The OpenAI migration guide in our docs maps OpenAI’s transcription and Realtime transcription settings to Cartesia’s.

Related guides

• Best speech-to-text models compared
• Deepgram pricing 2026
• ElevenLabs pricing 2026
• Word error rate, explained