Tool Suggestions
Mistral Voxtral 3: The Voice-Agent SDK That Acts During the Call — From $0.04 a Minute
Building a voice agent has usually meant three vendor contracts: one API to transcribe, a language model to reason, and another engine to speak. Mistral's Voxtral 3, launched October 20, collapses that into one product line — new speech-to-text and text-to-speech models plus a voice-agent SDK that listens, reasons with tool calls, and answers as a single streaming pipeline. The headline price is $0.04 per minute for the full agent when Mistral's own Small 4 model does the reasoning. The part European teams will notice first is where it can run: Mistral's European endpoints, or on your own infrastructure.
What shipped
The family has four pieces. Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3 handle speech-to-text across 50+ languages, with diarization and context biasing; the Mini model is 4 billion parameters. Voxtral Mini Transcribe Realtime 3 does live transcription at under 500 ms latency. Voxtral Speech 3 generates speech in 35+ languages and 50+ dialects, and its Pro Voice option clones a voice from 30 minutes of audio. The agent layer sits on top: an SDK built on LiveKit that connects over SIP for phone calls and WebRTC for browsers and apps, integrates with existing IVR and contact-center software, and — the substantive feature — can call functions, search the web and retrieve from your own systems in the middle of a conversation. Mistral's worked example is a patient giving a referral number; the agent retrieves the doctor's letter, translates it for the caller, and books the appointment, all inside one call. Each stage is a separate model, so the reasoning LLM, the voice or the language can be swapped without rewriting the agent. Mistral claims turns complete in under a second.
The pricing, in full
Transcription is priced per minute: $0.0035 for Mini batch, $0.0040 for Small, $0.008 for Mini realtime. Speech generation is $13 per million characters. The complete voice-agent service — transcription, reasoning and spoken response together — is $0.04 per minute with Mistral Small 4 as the reasoning model, and $0.08 per minute with any other reasoning model; at 10,000 minutes, that is $400 versus $800, before telephony and implementation costs. The per-minute agent rates are roughly half of what OpenAI's Realtime API charges for a comparable pipeline, which is plainly the comparison Mistral wants drawn.
The benchmark claims — vendor-reported
Mistral reports a macro word error rate of 4.1% (language specified) and 4.0% (auto-detected) for Voxtral Small Transcribe 3 across 45 FLEURS languages — tied with Microsoft's MAI-Transcribe-2 and ahead of ElevenLabs Scribe v2 at 4.9% and Whisper large-v3 by a wide margin — plus a diarization error rate of 18.4%, well below the previous Voxtral generation. Artificial Analysis' independent measurements put the model's accuracy in the same leading group (a 2.1% WER on its suite) though they rank its transcription speed below some rivals. Treat the launch comparisons as Mistral's case: the announcement names no independent reviewers or customers, some prices in its own materials are inconsistently presented, and the sub-second turn claim is stated without conditions. The FLEURS numbers are at least checkable against a public benchmark, which is more than most voice launches offer.
The deployment angle
Voxtral 3 is available through Mistral's European, US and global API endpoints, and OpenRouter lists the new models with EU-hosted, zero-data-retention endpoints. Full self-deployment on private infrastructure is offered through Mistral's sales team, with terms not published in the announcement. For European organizations that cannot send call audio to a shared US cloud — a real constraint in health, finance and public services — a credible voice stack with an EU home is the differentiator, more than the per-minute price. One compliance note travels with any EU voice deployment: the EU AI Act's transparency and audio-watermarking obligations apply to AI voice agents operating in the Union regardless of vendor, and they are the deployer's responsibility, not Mistral's.
The takeaway
Voxtral 3 is a serious one-vendor entry into voice agents: strong multilingual transcription at low prices, tool calling during live calls, and deployment options that US-first rivals do not match. The unverified parts are the usual ones — launch benchmarks, latency claims, unpublished self-hosting terms — and they are worth an afternoon's pilot before any commitment. For teams already comparing voice generators on output quality alone, this release widens the question: not just how the voice sounds, but where the call is allowed to live.