What is the cheapest Whisper API?

Because Whisper is open-weight, specialized inference hosts serve it well below OpenAI's own per-minute rate. The cheapest endpoint is rarely the official one; compare current per-minute prices in the Perkstack rankings.

Is there a free way to run Whisper?

Yes. Some inference hosts with free tiers serve Whisper, and you can self-host it on a free GPU notebook for experiments. For real volume, a cheap managed host is usually most cost-effective.

Should I use Whisper or another STT model?

Whisper is a strong default, but purpose-built APIs like Deepgram or AssemblyAI sometimes win on price or accuracy for a given language or mode. Compare them in our cheapest speech-to-text guide.

How do I lower transcription costs?

Use batch instead of streaming where latency allows, right-size the model, and route to the cheapest host per model, re-checking periodically as prices change.

Whisper API: The Cheapest and Free Ways to Run It in 2026

OpenAI's Whisper is open, so you rarely need to pay OpenAI's rate. Here are the cheapest Whisper API hosts in 2026, the free ways to run it, and when another STT model wins.

Whisper is still one of the most popular speech-to-text models, and because the weights are open, you almost never need to pay OpenAI's own per-minute rate to use it. Here is how to run the Whisper API cheaply or free in 2026, and when a different model is the better call.

Compare live per-minute prices in the rankings.

Why you should not default to OpenAI's Whisper rate

OpenAI hosts Whisper, but it is open-weight, so many providers serve the same model far below OpenAI's per-minute price. Specialized inference hosts run Whisper-large at a fraction of the cost, which is why the cheapest Whisper endpoint is rarely the official one. See the Whisper Large v3 tracker for the current spread.

Free and cheap ways to run Whisper

Free inference tiers. Some hosts with free tiers serve Whisper, so you can transcribe at no cost while prototyping.
Cheap specialized hosts. Providers that focus on fast inference run Whisper-large at very low per-minute rates; we track the cheapest in the rankings.
Self-host. Whisper runs locally or on a free GPU notebook for experiments. See where to get free GPU compute. For real volume, a managed host is usually cheaper once you count ops time.

When another STT model wins

Whisper is a strong default, but purpose-built speech-to-text APIs sometimes beat it on price or accuracy for a given language or mode. We track several alongside Whisper, including Deepgram Nova-3 and AssemblyAI. For a full comparison, see the cheapest speech-to-text API.

Keep transcription costs down

Use batch, not streaming, where latency does not matter; batch is usually cheaper.
Right-size the model. A smaller, faster variant is often accurate enough.
Compare hosts per model and re-check periodically, since prices move.

The bottom line

Running Whisper cheaply in 2026 is mostly about not paying the official rate: route to a cheap specialized host or a free tier, and compare against purpose-built STT models for your use case. See the rankings, grab voice credits in the catalog, and create a free account.