The product is the pipeline, not a model
Earlier voice launches sold a component: a better transcriber, a more natural voice. Voxtral 3 sells the assembled thing. The SDK keeps speech in, reasoning and speech out running as one streaming pipeline rather than a chain of separate calls. The realtime transcriber detects when a caller has finished speaking, a general model reasons and calls tools, and the speech model answers in a tone fitted to the conversation so far. During a call the agent can verify a caller's identity, check a bill or book an appointment, and it works over SIP phone lines, WebRTC in browsers and apps, and existing contact-centre software.
That packaging is the actual competitive move. ElevenLabs also announced Eleven Realtime v1 Max and Mini on 20 October. According to The Decoder, ElevenLabs reported a 98.8% Big Bench Audio result and a 0.57-second response figure for Max. Those are ElevenLabs' own measurements, not a controlled comparison with Voxtral 3. The same-day launches suggest competition is increasingly about the complete conversational experience — latency, interruptions, reasoning and speech — not transcription accuracy alone. Mistral's answer is modular underneath, since each stage is a separate model and the reasoning model can be swapped, but it is sold as one stack with one bill.
Four cents a minute, with the conditions attached
The pricing table is the strongest page in the announcement because anyone can check it. At Mistral's published agent API rate, $0.04 per minute works out to $2.40 for 60 minutes of billed conversation, with Mistral Small 4 providing the reasoning, and transcription and speech included. Choosing another reasoning model raises the published rate to $0.08 per minute, or $4.80 for 60 minutes. These are calculations from Mistral's rate card, not measured total operating costs for a production deployment. And the cheap transcription headline, $0.0035 a minute, is the batch price for the small 4B model. The realtime transcriber an agent actually uses costs $0.008 a minute on its own, which is still below one cent per minute, but it is more than twice the headline figure.
Speech output is priced by character, at $13 per million characters for both batch and realtime. Mistral says that undercuts OpenAI's TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. That comparison is Mistral's, drawn from vendor pricing pages, and it is the kind of claim a buyer can verify in an afternoon. The launch would be stronger if more of it worked that way.
The benchmarks, in small print
Mistral states plainly that Small Transcribe 3 achieves state-of-the-art results in diarization and multilingual transcription and outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe. The supporting charts are real benchmarks with named datasets: a macro-average word error rate across 45 languages from FLEURS, diarization scored on AMI, Switchboard and AISHELL-4, keyphrase recall on Earnings-22, and an Artificial Analysis index built from VoxPopuli, Earnings-22 and AgentTalk. The transcription comparisons use named datasets and stated metrics, giving outsiders a clearer basis for reproduction.
The speech side is weaker evidence. Several of the text-to-speech evaluations are labelled, in Mistral's own footnotes, as internal evals, run on a few hundred prompts selected for release, and scored by Gemini 3.8 Flash, a competitor's model. Using a competitor's model as judge does not make the result independent: the scores can still depend on the judge's preferences, on the prompts Mistral chose, and on a one-to-five scale with no published raw outputs. These are Mistral-reported results, not independently reproduced performance findings in the sources reviewed. The announcement's own footnotes are what tell you so.
One more distinction is worth keeping. The "50+ languages" in the headline is the offline Small transcriber's count. The realtime model covers more than 35 languages, and speech covers more than 35 languages with 50-plus dialects. A team building a live agent in a less common language is buying the smaller number, not the headline one.
Sovereignty, arranged by sales call
Mistral's pitch leans on control: run the stack on your own infrastructure, keep sensitive conversations in your environment, decide whether audio is retained. For European buyers and regulated industries that is a genuine difference from a US-hosted API, and it is the part of the launch aimed past OpenAI and ElevenLabs.
Read the offer closely, though. The announcement first describes availability through Mistral's EU-hosted API, then, in the closing section, through European, US and global endpoints. Self-deployment could give customers meaningful control over where inference runs and how audio is retained. But Mistral's announcement does not offer the full stack as a public weights download: prospective customers are directed to its team. That makes deployment terms, infrastructure requirements, data handling and total cost questions to resolve before purchase, rather than capabilities a reader can assume from the API launch alone.
Sources
- [1] Mistral AI, 'Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages', 20 October 2026Read source
- [2] The Decoder, 'Mistral AI launches Voxtral 3, a full voice agent stack that listens, thinks, and talks in real time', 20 October 2026Read source
- [3] The Decoder, 'ElevenLabs releases two new speech-to-speech models designed for real-time AI conversations', 20 October 2026Read source