The Voice Agent Now Bills by the Minute

## Four cents buys the whole call — listening, thinking, and talking included

The interesting thing about Mistral’s Voxtral 3 is not that a model learned to listen in fifty languages. It is that the voice agent stopped being three separate bills.

Until now, a spoken agent was an assembly job. One vendor transcribed the caller, a second model reasoned about what to say, a third vendor spoke the answer back, and the builder owned every seam between them — the latency, the dropped context, the three invoices. Voxtral 3, released by Mistral on October 20, collapses that into one streaming pipeline and one number: the voice agent API costs $0.04 per minute when Mistral Small 4 does the reasoning, and $0.08 per minute with a larger reasoning model, according to the price table in Mistral’s own announcement. The Hacker News discussion, where the launch sat near the top of the front page with 636 points at fetch time, circled the same point from the other side — the models are good, but the packaging is the story.

That sounds like a bundling discount. It is more than that. It changes what the unit of cost is. A voice agent is no longer priced in tokens you cannot predict. It is priced in minutes you can.

The price sheet, line by line

Mistral’s announcement lists every stage separately, which is what makes the bundle readable. Batch transcription with Voxtral Mini Transcribe 3 costs $0.0035 per minute. The flagship Voxtral Small Transcribe 3 costs $0.0040 per minute. The realtime transcription model costs $0.008 per minute. Speech generation, Voxtral Mini TTS 3, costs $13 per million characters. And the assembled agent — transcription, reasoning, tool calls, and speech in one pipeline — costs $0.04 or $0.08 per minute depending on the reasoning model.

Do the arithmetic the way a builder has to. Transcribing an hour of audio on the Mini model costs $0.21 at Mistral’s listed price. A ten-minute support call run end to end through the agent API costs $0.40 with the smaller reasoning model. The reasoning choice alone doubles the minute rate, which means the most expensive decision in a voice agent is no longer the speech. It is how smart you want the middle of the call to be.

That is the honest shape of this release. Speech got cheap enough to itemize. Thinking is still the line that moves.

What the models actually claim

The transcription flagship, Voxtral Small Transcribe 3, covers 50 languages including the 24 official EU languages, with speaker diarization, context biasing for domain vocabulary, and end-of-turn detection, per Mistral’s announcement. Mistral reports a macro-average word error rate of 4.1% with the language specified and 4.0% with it auto-detected across 45 languages on the FLEURS benchmark, ahead of Whisper large-v3 and Gemini 3.5 Transcribe in its own comparison table. Artificial Analysis, in coverage of the launch, puts the Small model at a 2.1% weighted word error rate on its own index — a different test with a different weighting, and a reminder that every vendor’s favorite number depends on which benchmark is doing the weighing.

The Mini transcription model is 4 billion parameters, per Mistral, aimed at cheaper and faster batch work. The speech model supports more than 35 languages and over 50 dialects at the $13 per million characters price, which Mistral says undercuts OpenAI’s TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash. Pro Voice adapts a speaker’s voice and style across languages from 30 minutes of audio. These are vendor claims, presented on a vendor page, and the launch coverage is right to treat the comparisons as Mistral’s until independent testing catches up.

The catch is a dial, not a defect

Realtime is where the trade lives. Mistral says its realtime transcription model answers at sub-500 millisecond latency in over 35 languages, and that a full agent turn — listen, reason, speak — completes in under one second. But the company’s own FLEURS 35 comparison, as reported in the launch coverage, shows what speed costs: at a 240 millisecond delay setting, the realtime model’s word error rate is 7.4%, against 4.2% for Gemini 3.5 Realtime in the same table. Give the system more time and the errors fall. Take the time away and they climb.

That is not a scandal. It is the physics of streaming speech, printed in the vendor’s own chart. But it means the headline number and the usable number are different settings on the same dial. A builder quoting sub-500 millisecond latency to a customer is also, quietly, accepting roughly three extra transcription errors in every hundred words at the fastest setting — on a clean benchmark, before a real caller with a bad connection and an accent the training data rarely met starts talking.

The second catch is quieter. Voxtral 3 is available through Mistral’s API on European, US, and global endpoints today. Running the full stack on your own infrastructure is offered, but the announcement routes it through a sales conversation — contact the team for on-premises or private cloud deployment. The coverage of the launch notes that no public downloadable weights for Voxtral 3 have been established. For a company whose reputation was built on open weights, this generation of its speech stack is, for now, an API product with a self-hosting conversation attached. The sovereignty pitch is real — EU hosting, control over whether audio is retained — but sovereignty by contract is a different thing from sovereignty by download.

Why the packaging still matters

None of that erases the structural point. The voice agent SDK is built on LiveKit, takes calls over SIP and sessions over WebRTC, connects to existing telephony and contact-center software, and keeps each stage as a separate model underneath — Mistral’s own line is that you can change the LLM, the voice, or the language without rewriting the agent. The stages are separable. The bill is not. That combination is the product: modular parts, unified meter.

Hacker News commenters have spent two years watching agent demos that worked in a recording and died in a call queue. A per-minute price does not fix reliability, but it does something demos never do — it makes the cost of failure legible. A botched ten-minute call costs forty cents of pipeline and whatever the customer’s patience was worth. The second number is still the bigger one. It always was.

For builders, the takeaway is practical: price your voice agent in minutes, not tokens, and test it at the latency setting you actually plan to ship, not the one in the launch chart. The transcription is cheap, the speech is cheap, and the reasoning in the middle is priced like it knows it is the product.

The meter was always running on voice. Now it fits on one line of a price table.

Source: Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages