Yes. The weights are open and there is no licence fee. That is the answer, and it is the least useful true thing about Whisper's cost.
Free weights are not a free service. You pay for the hardware that runs them, and for a voice agent you pay for something Whisper does not include at all.
Whisper was not built for streaming
A voice agent needs partial results while the caller is still speaking, so it can detect when they have finished a turn. Whisper transcribes a finished audio file. Those are different jobs.
Bridging the gap means a streaming wrapper, turn detection and reconnection handling, all of which somebody has to maintain. By the time that exists, the commercial streaming vendors are usually cheaper than the engineer maintaining it, and that is the arithmetic worth doing rather than the licence one.
What recognition costs when you buy it
| Option | Published rate | Streaming? |
|---|---|---|
| Whisper, self-run | no licence fee, compute is yours | Not natively |
| Deepgram streaming list | $0.0077/min | Yes, with turn detection |
| Deepgram, promotional column | $0.0065/min | Yes |
Recognition is roughly a sixth of the $0.0465 component floor for a voice agent minute. It is not where the money is. Synthesis, which runs from about $0.014 to $0.085 a minute depending on the voice, moves a bill further than any recognition decision will.
When self-running actually wins
- Batch transcription of recorded files, where latency does not matter and nobody is waiting on a turn.
- Volume high enough that owning the compute beats a per-minute rate, which is a real threshold and one you can compute from your own minutes.
- Data residency or privacy requirements that rule out sending audio to a third party at all.
- You already run GPUs for something else and the marginal cost of another workload is close to zero.
Outside those four, a managed streaming service is usually the cheaper answer once you price the engineering time honestly. The mistake is comparing a licence fee of zero against a rate of $0.0077 and stopping there.
What we do not publish
We do not compare recognition vendors on accuracy. We have not measured word error rate, and every published benchmark we found was run by a party selling one of the models in it. What we verify is price, and on price the spread is large, checkable and dated.