Mistral Voxtral 3 Explained: Voice Agent SDK, Pricing, and Tool Calling

Mistral dropped Voxtral 3 on October 20, 2026: a new family of speech-to-text and text-to-speech models plus a voice agent SDK that bundles listening, reasoning, tool calling, and speaking into one real-time pipeline. You can call it through Mistral's API starting today, and the headline number is the one that matters most if you're building anything that picks up a phone: a full voice agent conversation costs $0.04 per minute when Mistral Small 4 handles the reasoning step.

That's the pitch in one sentence. Here's what's actually in the box, what it costs, how the agent SDK is wired together, and where I'd be careful before shipping it.

What is Voxtral 3?

Voxtral 3 isn't one model, it's four, plus an SDK that glues them together:

• Voxtral Small Transcribe 3 — the flagship speech-to-text model, strongest accuracy across 50+ languages
• Voxtral Mini Transcribe 3 — a smaller, cheaper transcription model for batch jobs
• Voxtral Mini Transcribe Realtime 3 — the streaming variant built for sub-second latency during live calls
• Voxtral Speech 3 (also shown as Voxtral Mini TTS 3.0) — text-to-speech with voice cloning and emotional control

On top of those four sits the Voice Agent SDK, built on LiveKit, which runs transcription, an LLM, and speech synthesis as one streaming loop instead of three separate API calls you have to stitch together yourself.

This is Mistral's second Voxtral generation. The first Voxtral, released back in mid-2025, shipped as open weights. Voxtral 3 does not. I'll get to why that matters in a minute, because it's the thing Hacker News latched onto fastest.

How much does Voxtral 3 cost?

Pricing splits into two buckets: pay per model, or pay per agent-minute for the whole pipeline.

The agent price covers the whole round trip: speech in, a reasoning model thinking (and calling tools if it needs to), speech back out. That's a genuinely aggressive number. A 10-minute support call on the cheap tier runs 40 cents. Swap in a bigger reasoning model and it doubles to 80 cents, still cheap compared to running a human agent, or honestly compared to stitching together Whisper plus GPT-4 plus ElevenLabs yourself and paying each vendor separately.

How accurate is it, actually?

Mistral published FLEURS benchmarks across 45 languages: Voxtral Small Transcribe 3 lands at 4.1% word error rate with a forced language setting and 4.0% when the model infers the language itself. Voxtral Mini comes in at 4.6%.

Artificial Analysis ran its own independent AA-WER Index, which weights three real-world datasets (agent calls, VoxPopuli audio, earnings call transcripts) instead of the cleaner FLEURS set, and the results hold up:

• Voxtral Small Transcribe 3: #3 overall on the Speech-to-Text leaderboard at 2.1% AA-WER, ahead of ElevenLabs Scribe v2 (2.2%) and Grok Voice Transcribe 2.0 (2.3%)
• Voxtral Mini Transcribe 3: #5 at 2.2%
• Voxtral Mini Transcribe Realtime 3: 3.3% AA-WER at a 960ms latency setting, #6 of 46 streaming models tested, and 3.6% if you push latency down to 480ms

For diarization (telling speakers apart on a call), Small scores 18.4% and Mini 18.6% on concatenated minimum-permutation WER, both beating Pyannote Orchestrator and Microsoft's AI Transcribe 2. On the TTS side, Mistral claims it beats ElevenLabs Flash, Deepgram Flux, OpenAI's TTS-1, and Gemini Flash on a quality-to-cost basis, and Artificial Analysis has it entering the top 12 on both of its Text-to-Speech arenas. I haven't run my own side-by-side yet, so take Mistral's framing with the usual grain of salt a vendor's own comparison deserves, but the AA-WER numbers are third-party and they're genuinely good, not just "good for the price."

How does the Voice Agent SDK actually work?

This is the part that's new, not just "another Whisper competitor." The SDK runs three stages as one streaming pipeline instead of a request-response chain:

1. Voxtral Mini Transcribe Realtime 3 listens continuously and detects when the caller stops talking
2. A reasoning model (Mistral Small 4 by default) decides what to say back, and can call tools mid-turn: web search, your document library, workspace connectors, or your own functions
3. Voxtral Speech 3 generates the spoken reply in a tone that matches the conversation so far

Each turn is designed to complete in under a second, and the agent can be interrupted mid-sentence if the caller starts talking over it, which is the detail that actually makes a voice bot feel less robotic on a real call. It handles phone calls over SIP and browser/app sessions over WebRTC, and because each stage is a swappable model, you can change the LLM or the voice without rebuilding the pipeline.

Here's roughly what wiring a tool-calling agent looks like using LiveKit's Mistral plugin:

from livekit.agents import AgentSession, function_tool
from livekit.plugins import mistralai

@function_tool
async def check_order_status(order_id: str) -> str:
    # your own lookup logic runs here
    return f"Order {order_id} shipped yesterday, arriving Thursday."

session = AgentSession(
    stt=mistralai.STT(model="voxtral-mini-transcribe-realtime-3"),
    llm=mistralai.LLM(model="mistral-small-4", tools=[check_order_status]),
    tts=mistralai.TTS(model="voxtral-mini-tts-3"),
)

If you'd rather stay inside Mistral's own SDK instead of LiveKit's wrapper, the realtime client ships as an extra on the main Python package: pip install "mistralai[realtime]".

Is Voxtral 3 open source?

No, and this is where I part ways with the launch-day hype a little. The original Voxtral was open weights. Voxtral 3 is not — it's API-only, or "self-deployed" under an enterprise agreement on your own infrastructure, which is a very different thing from downloading a checkpoint and running it on your laptop. The Hacker News thread on this called it out within the first few comments, and I think they're right to. If you specifically picked Mistral in the past because you could self-host for privacy or cost reasons, Voxtral 3 doesn't give you that anymore. You're back to trusting a hosted API, same as OpenAI or ElevenLabs, just a cheaper one.

There's also a fair complaint buried in that same thread about European language coverage being thinner than you'd expect from a French company selling data-sovereignty as a feature. Worth checking your specific language before you commit, not just trusting the "50+ languages" headline number.

Where does Voxtral 3 land vs the competition?

The thing nobody else is really offering is the bundled, priced-per-minute agent pipeline. You can build the same thing yourself with Whisper plus any LLM plus ElevenLabs, and plenty of teams do, but you're paying three vendors and gluing three APIs together with your own latency budget. Mistral's pitch is that you get one number and one SDK instead.

My take

If you're already in the Mistral ecosystem, or you need EU-hosted audio processing for compliance reasons, this is an easy yes — the pricing is hard to beat and the benchmarks aren't vendor-cherry-picked, Artificial Analysis backs most of it independently. If what you actually wanted was an open-weights model you could run on your own GPU for free, this launch is a step backward from the original Voxtral, and I wouldn't pretend otherwise. Test it on your worst real calls before you trust it in production: background noise, accents, people talking over the bot. The demos are always clean. Real call centers aren't.

If you're also tracking what else Mistral shipped this month, we covered their CPU-friendly forecasting model in our Timestral walkthrough, and the broader Microsoft deal that's funding this release pace in our Mistral-Microsoft partnership piece. For how Voxtral 3 stacks up against other speech launches this quarter, see our DeepInfer breakdown.

Watch and read more

FAQ

What is Mistral Voxtral 3?

Voxtral 3 is Mistral AI's latest speech lineup, released October 20, 2026: three speech-to-text models (Small, Mini, and Mini Realtime), one text-to-speech model, and a LiveKit-based Voice Agent SDK that combines transcription, an LLM, and speech synthesis into one real-time, tool-calling pipeline.

How much does Voxtral 3 cost?

Transcription starts at $0.0035 per minute for batch jobs and $0.008 per minute for realtime streaming. Text-to-speech is $13 per million characters. A full voice agent conversation costs $0.04 per minute with Mistral Small 4 as the reasoning model, or $0.08 per minute with other reasoning models.

Is Voxtral 3 open source?

No. The original Voxtral released in 2025 shipped as open weights, but Voxtral 3 is closed and API-only. "Self-deployment" means running Mistral's software on your own infrastructure under an enterprise agreement, not downloading a model checkpoint to run locally.

How accurate is Voxtral 3 compared to Whisper and ElevenLabs?

On Artificial Analysis's independent AA-WER Index, Voxtral Small Transcribe 3 ranks #3 overall at 2.1% word error rate, ahead of ElevenLabs Scribe v2 (2.2%) and Grok Voice Transcribe 2.0 (2.3%). Voxtral Mini Transcribe 3 ranks #5 at 2.2%.

Can I add tool calling to a Voxtral 3 voice agent?

Yes. The Voice Agent SDK supports built-in tools (web search, document libraries, workspace connectors) and your own custom functions, which the reasoning model can call mid-conversation before speaking its response back to the caller.

What languages does Voxtral 3 support?

Transcription covers 50+ languages including all 24 official EU languages plus several Asian and Indic languages. The text-to-speech model covers 35+ languages across more than 50 dialects.

Does Voxtral 3 work over phone calls?

Yes. The Voice Agent SDK supports SIP for traditional phone integration and WebRTC for browser or app-based sessions, with each conversational turn designed to complete in under one second.