What Is Voxtral 3? Pricing, Voice Agents & Proven Uses

A phone agent that listens, checks your booking system, and answers out loud now costs about four cents a minute. That’s the headline number behind Voxtral 3, which Mistral AI released on 20 October 2026.

So what is Voxtral 3? It’s Mistral’s family of speech models plus a voice agent SDK. Three models turn speech into text. One model turns text into natural speech. The SDK joins them with a reasoning model, so you can build voice agents that talk, think, and act on business tools in real time.

In this guide, you’ll get the full price list, a monthly budget at real call volumes, and a fair comparison with rivals. We also cover the limits that were often overlooked in most launch coverage.

Key Takeaways

• Mistral’s new stack launched on 20 October 2026 with three transcription models, one text-to-speech model, and a voice agent SDK built on LiveKit.
• Transcription starts at $0.0035 a minute. A full voice agent costs $0.04 a minute with Mistral Small 4.
• Batch transcription covers 50 languages. Speech output covers 35+ languages and 50+ dialects.
• Phone calls are enterprise-only and inbound-only for now, and agents can’t hand calls to a human.
• The benchmark wins are Mistral’s own tests on clean speech. Test your real audio first.

What Is Voxtral 3? The Models in the Family

Voxtral 3 is not one model. It’s a set of speech tools you can use alone or join together. Mistral first entered voice in July 2025. Then it added Voxtral Transcribe 2 in February 2026 and its first text-to-speech model in March 2026, according to TechTimes. Voxtral 3 pulls all of that into one stack.

Here’s what ships in the family, based on Mistral’s launch post:

Model	What it does	Key detail
Voxtral Small Transcribe 3	Top-accuracy transcription	Labels who said what; takes a list of key terms
Voxtral Mini Transcribe 3	Cheaper batch transcription	4B parameters
Voxtral Mini Transcribe Realtime 3	Live transcription for agents	Under 500 ms; knows when a caller stops talking
Voxtral Mini TTS 3	Text-to-speech	Emotion, pace, and emphasis control; voice cloning
Voice agent SDK	Joins it all into one pipeline	Built on LiveKit

Two features matter most for business use. First, diarization tells you who spoke when. That’s vital for sales calls and meetings. Second, context biasing lets you hand the model a list of product names or jargon. It then spells those words right without any extra training.

Language reach is wide. The batch models handle 50 languages, including all 24 official EU languages. The live model covers 35+ languages. The speech model covers 35+ languages and 50+ dialects.

There’s also Pro Voice. It builds a custom voice from about 30 minutes of audio. That’s aimed at brands that want one voice across many languages.

What most guides miss: the parts are swappable. Each stage (listen, think, speak) is a separate model. So you can keep your own speech-to-text vendor and use only Voxtral’s voice. Or you can run Voxtral transcription with a different language model. That flexibility lowers your lock-in risk.

How Voxtral 3 Voice Agents Work

A Voxtral 3 voice agent runs three steps as one live stream: it listens, it reasons, and it speaks. Each turn in a conversation finishes in under one second, Mistral says.

Here’s a simplified explanation of the flow:

1. Listen. The real-time model transcribes the caller as they talk. It also detects when they finish speaking.
2. Think. A reasoning model, such as Mistral Small 4, decides what to do. It can call tools, search the web, or pull answers from your systems.
3. Speak. The speech model replies out loud. It matches the tone of the call so far.

If the caller cuts in, the agent stops talking. That sounds small. However, it’s the single biggest reason older voice bots felt robotic.

The SDK connects in two ways. Agents can answer and place calls over SIP, the standard for internet phone lines. They can also run inside browsers and apps over WebRTC. Mistral says the SDK also plugs into IVR systems and contact-center software.

Now for the catch. Phone calling isn’t self-serve yet. Eesel’s pricing analysis quotes Mistral’s docs say the phone feature is for enterprise customers only. The same docs say the phone agent handles inbound calls only and can’t transfer a call to a person.

For a startup, that changes the plan. If you need a human handoff for angry customers, you can’t do that over the phone line today. So start with an in-app or web voice assistant, where the self-serve SDK already works.

Then move to phone lines once you’re on an enterprise deal or Mistral opens access.

Voxtral 3 Pricing: What a Minute Really Costs

Voxtral 3 pricing runs from $0.0035 a minute for batch transcription to $0.08 a minute for a full voice agent with a larger model. All prices below are in USD and come from Mistral’s October 2026 launch post.

Product	Price
Voxtral Mini Transcribe 3 (batch)	$0.0035 per minute
Voxtral Small Transcribe 3	$0.004 per minute
Voxtral Mini Transcribe Realtime 3	$0.008 per minute
Voxtral Mini TTS 3	$13 per 1 million characters
Voice agent API with Mistral Small 4	$0.04 per minute
Voice agent API with a larger model	$0.08 per minute

Let’s turn that into a budget. One hour of recorded audio through Voxtral Mini Transcribe 3 costs $0.21. A small clinic taking 3,000 agent minutes a month would pay $120 with Mistral Small 4. At 10,000 minutes, the bill is $400, or $800 with a larger model, as Search Engine Watch worked out.

Here’s an angle few reviews cover. You can build the same agent from the parts and pay much less. Eesel, which is built in one minute from separate Voxtral pieces, costs about $0.015.

That’s live transcription at $0.008, speech at roughly $0.005, and model tokens at about $0.001. The $0.04 bundle costs about 2.7 times more.

So which should you pick? The extra $0.025 a minute pays for turn-taking, interruptions, audio transport, and tool wiring. For most early-stage teams, that’s a fair trade. Building turn detection yourself can take weeks of engineering time.

But once you pass tens of thousands of minutes a month, the DIY route starts to pay off.

Keep three hidden costs in mind:

• Phone lines are priced by contract, not on the public sheet.
• Tool calls are billed per call on top of the minute rate.
• On-premises hosting needs a quote from Mistral’s sales team.

Furthermore, treat these as launch prices. MarkTechPost reported that all four models launched in public preview, so rates may shift. Check Mistral’s pricing page before you lock a budget.

How Voxtral 3 Compares With OpenAI, ElevenLabs, and Deepgram

On price, Voxtral 3 is the cheapest bundled voice agent among the major vendors right now. On accuracy, it leads Mistral’s own charts. But those charts deserve a careful read.

Here’s how the per-minute agent rates stack up, using eesel’s October 2026 comparison at 20,000 minutes a month:

Voice agent	Per minute	Monthly (20,000 min)
Voxtral 3 with Mistral Small 4	$0.04	$800
OpenAI GPT-Live-1	$0.05	$1,000
Cartesia agents	$0.06	$1,200
Deepgram Voice Agent (Standard)	$0.075	$1,500
ElevenLabs ElevenAgents	$0.08	$1,600

However, these rates don’t cover the same things. Some rivals include phone numbers on self-serve plans. Others charge for the language model on top. Compare full bills, not just the headline rate.

On accuracy, Mistral reports a word error rate of about 4.1% for Voxtral Small Transcribe 3 on the FLEURS test. Its chart puts ElevenLabs Scribe v2 at 4.9%, and GPT Transcribe at 6.3%. Lower is better. Voxtral also scored about 94.6% recall on key phrases when given context, up from 81.4% without it.

That said, there’s a big caveat. FLEURS uses clear, readable speech. TechTimes notes that real phone audio, with noise and accents, often produces two to four times higher error rates. No independent tests on live calls were public at launch.

Our take: Mistral’s stack wins clearly on price and on European hosting. On accuracy, it’s a strong contender, not a proven winner. We recommend running 50 to 100 of your actual calls through two vendors before making a commitment.

Proven Business Use Cases for Founders

This stack fits best where calls are frequent, repetitive, and tied to a business system. Mistral names contact centers, service desks, and hands-free assistants as core targets. No named customer case studies were public at launch. So the uses below are proven categories for voice AI, matched to what Voxtral 3 can do today.

Customer Support and Service Desks

Most support calls ask the same few things: order status, password resets, and bill checks. A voice agent can answer these and look up records through tool calls. As a result, your human team handles only the challenging cases.

See our guide to AI customer service tools for startups. Just remember the current phone limit: no transfer to a person. So keep a clear “call us back” path for edge cases.

Appointment Booking and Reminders

Clinics, salons, and repair shops lose bookings when nobody picks up. Mistral lists appointment booking as a core use. The agent checks your calendar, offers slots, and confirms the time out loud. For a small business, that’s a 24/7 front desk at a few cents a minute.

Meeting and Sales Call Transcription

You don’t need the full agent to get value. Voxtral Small Transcribe 3 labels each speaker and spells product names right with context biasing. For example, a B2B startup can transcribe every sales call, then feed the text to a CRM.

Pair it with the business automation tools we’ve tested. At $0.004 a minute, 500 hours of calls cost $120.

Multilingual Products for Global Markets

If you sell across Europe or Asia, language coverage matters more than raw accuracy in English. Fifty transcription languages and 50+ speech dialects let one agent serve many markets.

Additionally, Mistral provides hosting on European, US, and global endpoints. That helps founders selling to EU clients who ask where their audio data is stored.

Limits and Risks to Check Before You Build

Voxtral 3 is cheap and capable, but four gaps could slow a launch. Check each one against your plan.

1. No download yet: You can’t grab the model weights and run them yourself. As of 5 November 2026, ReflektLab found no Voxtral 3 weights on Mistral’s Hugging Face page. Self-hosting goes through Mistral’s sales team.

2. Phone features are limited: As covered above, phone calls are enterprise-only and inbound-only and can’t be transferred to a human.

3. EU AI Act duties apply: Under Article 50, a voice agent talking to people must say it’s an AI at the start of the call. AI-made audio also needs a machine-readable marker. TechTimes reports that the Voxtral 3 SDK does not support either requirement by default. So your team must build both. Fines can reach €15 million or 3% of global turnover.

4. Voice cloning needs consent: Pro Voice can copy a voice from 30 minutes of audio. That’s useful for brands. But it’s also a fraud risk. Mistral hadn’t published a detailed consent policy at launch.

Obtain written consent from anyone whose voice you clone. Furthermore, the US FCC banned AI-generated voices in robocalls in February 2024 (an older rule, still in force).

A common mistake is to treat these as “later” problems. In fact, the disclosure line costs one sentence of script. Add it on day one.

Conclusion

Voxtral 3 gives founders a full voice stack at a price that was difficult to imagine two years ago. Four cents a minute buys an agent that listens, reasons, uses your tools, and answers out loud.

The value is real, but so are the gaps. Self-serve access works today for web and app voice, while phone lines still need an enterprise deal. Mistral’s accuracy claims look strong, yet they come from clean test audio.

And EU rules on disclosure and audio marking are now your responsibility, rather than the SDK’s.

Here’s our advice. Pick one narrow, high-volume task, such as booking calls or support FAQs. Build it as an in-app voice assistant first. Then test 100 real conversations and compare the error rate and cost with one rival.

Have you tried Voxtral 3 or another voice agent in your startup? Share what worked, or read our guide to the best AI voice generators.

FAQs

1. How much does Voxtral 3 cost per minute?

It costs $0.0035 a minute for batch transcription, $0.004 for the top transcription model, and $0.008 for live transcription. A full voice agent costs $0.04 a minute with Mistral Small 4, or $0.08 with a larger model. Speech output is $13 per million characters.

2. Is Voxtral 3 open source or free to download?

No, not yet. As of early November 2026, Voxtral 3 weights weren’t on Mistral’s Hugging Face page. You can use it through the Mistral API or ask Mistral’s sales team about running it on your servers.

3. What languages does Voxtral 3 support?

Its batch transcription supports 50 languages, including all 24 official EU languages. Live transcription covers 35+ languages. The text-to-speech model speaks 35+ languages and 50+ dialects, with preset voices in 18 languages.

4. Is Voxtral 3 better than ElevenLabs for voice agents?

It’s cheaper, at $0.04 a minute versus $0.08 for ElevenLabs. Mistral’s tests also show lower transcription error rates. However, some rivals offer phone numbers on self-serve plans, which Voxtral doesn’t yet. So the better pick depends on your channel.

5. When was Mistral Voxtral 3 released?

Mistral AI released Voxtral 3 on 20 October 2026. The launch included three transcription models, one text-to-speech model, and a LiveKit-based voice agent SDK. It followed Voxtral’s first release in July 2025 and Voxtral TTS in March 2026.