Mistral Voxtral 3 Review: Speech Pricing and Voice Agent Claims Checked
Mistral launched three transcription models, a text-to-speech model and a voice agent SDK. We check the prices, the word error rates and the open-weights question.
Jump to 10 sections
Mistral launched Voxtral 3 on October 20, 2026, a speech stack with three transcription models, a text-to-speech model and a voice agent SDK. Batch transcription starts at $0.0035 per minute, and speech synthesis costs $13 per million characters. An independent leaderboard ranks the Small model third for accuracy, though open weights are not confirmed.
This review is for developers and product teams who build voice features such as call transcription, meeting notes or phone agents. It covers what shipped, what it costs, which results are independent and which are vendor claims. Model Drop reviews launches by checking each headline number against its source, and Voxtral 3 has a mix of both.
What Did Mistral Release With Voxtral 3?
Voxtral 3 is a speech stack, not a single model. According to OrcaRouter's breakdown, it includes Voxtral Small Transcribe 3, Voxtral Mini Transcribe 3 and Voxtral Mini Transcribe Realtime 3, plus a text-to-speech model and a voice agent SDK built on LiveKit, as summarized by OrcaRouter.
The SDK supports SIP, WebRTC, IVR and contact-center integrations. Mistral claims a full agent turn in under one second, with handling for barge-in, which means a caller can interrupt the voice agent mid-sentence.
Mistral's first speech-understanding models launched in July 2025, with text-to-speech enhancements in March 2026, according to Crypto Briefing's coverage of the new launch. Voxtral 3 is the first time the pieces arrive as one pipeline.
How Much Does Voxtral 3 Cost?
Prices are low. Voxtral Mini Transcribe 3 costs $0.0035 per minute, Voxtral Small Transcribe 3 costs $0.004 per minute, and the realtime model costs $0.008 per minute. Speech synthesis costs $13 per million characters. The voice agent SDK costs $0.04 per minute with Mistral Small 4 as the reasoning model and $0.08 per minute with another one.
Here is what that means in practice. A team that transcribes 1,000 hours of audio is billing 60,000 minutes. That costs $210 with Mini, $240 with Small and $480 with the realtime model.
For synthesis, OrcaRouter estimates about $0.70 per hour of speech, assuming roughly 900 characters per spoken minute. The table below compares the main prices.
| Component | Price | Example cost |
|---|---|---|
| Mini Transcribe 3 | $0.0035 per minute | $210 per 1,000 hours |
| Small Transcribe 3 | $0.004 per minute | $240 per 1,000 hours |
| Realtime Transcribe 3 | $0.008 per minute | $480 per 1,000 hours |
| Text-to-speech | $13 per million characters | About $0.70 per hour of speech |
| Voice agent SDK | $0.04 to $0.08 per minute | $2.40 to $4.80 per hour |
One outlet reported API access at about $0.001 per minute, but no other source we found repeats that figure, so treat it as unconfirmed. Our guide to text-to-speech models compared shows how these rates sit against other providers.
Which Voxtral 3 Results Are Independent?
A few are. Artificial Analysis, an independent tracker, put Voxtral Small Transcribe 3 at a 2.1% word error rate and third on its speech-to-text board. Mini scored 2.2%. Word error rate is the share of words a model gets wrong, so lower is better.
For context, the same breakdown lists Fun-Realtime-ASR-preview at 1.7%, Microsoft's MAI-Transcribe-2 at 2.0% and ElevenLabs Scribe v2 at 2.2%. Voxtral Small sits in a tight pack, and differences of a tenth of a point are small.
Speed differs more. The breakdown lists a median speed factor of 66.6 for Small, against 308.8 for MAI-Transcribe-2 and 20.3 for Fun-Realtime-ASR-preview. If throughput matters for you, the accuracy rank alone is not enough.
Which Claims Are Vendor-Reported Only?
Several. Mistral reports a 4.1% word error rate on FLEURS 45 with forced language and 4.0% with inferred language. It reports a diarization error of 18.4% and keyphrase recall rising from 81.4% to 94.6% with context biasing. Diarization means labeling who said what.
Mistral also says Voxtral 3 outperforms closed models from OpenAI and ElevenLabs on various benchmarks, and that it beats Whisper large-v3, according to the Crypto Briefing report. That article gives no specific scores, so the claim cannot be checked from the coverage.
The realtime model's sub-500 millisecond latency is also a claim. A phone agent feels natural only if each step is fast, so test it on your own network and audio.
How Good Is the Text-to-Speech Model?
Early signals are mixed. OrcaRouter reports that the model ranked 13th on the Provider Voice arena with 1,217 Elo, and 14th on Controlled Voice with 1,066 Elo, two days after launch. Rank bands are tied, and the sample is small.
Price helps. At $13 per million characters it sits below Cartesia Sonic 3.6 at $49 and ElevenLabs Eleven v4 at $80, which score 1,289 and 1,328 Elo. Cheaper models in the list, such as Simba 3.2 at $6.60, score 1,245.
The question is value. If your use case is high-volume narration or support calls, the price gap matters more than a 100-point Elo gap. If you need the most natural voice for a flagship product, the top-ranked models still lead. Our look at Grok Voice 3.0 and real call costs explains why total call cost includes more than the voice line.
What Do the Language Numbers Really Say?
They conflict. Mistral's summary cites more than 50 languages, while a more detailed breakdown cites more than 35 languages and more than 50 dialects. The realtime model's page gives both 35-plus and 40-plus languages in different places.
That matters if you serve a language at the edge of the list. Do not rely on a headline count. Test the exact language and accent you need, and ask Mistral which models support it.
A fair note in Mistral's favor is that it is a European lab and has focused on multilingual work. Still, the discrepancy is an easy thing to fix in documentation, and buyers should ask for clarity.
Is Voxtral 3 Open Weight?
Unclear. Mistral has leaned on open-weight models, and Crypto Briefing suggests developers can access the weights, without stating a license. OrcaRouter says open weights are not stated and the release does not appear to be downloadable. It notes that the March 2026 Voxtral TTS was released on Hugging Face under CC BY-NC 4.0, and that this release does not repeat that.
Access is through Mistral's API and Studio playgrounds, and through self-deployment on private infrastructure. Self-deployment is not the same as downloadable weights. It usually means a licensed deployment on your servers.
For privacy-sensitive audio, such as medical or legal calls, that distinction matters. Ask Mistral in writing what self-deployment includes and under what license.
Who Benefits Most From a Bundled Speech Stack?
Small teams benefit most. A startup building a phone assistant usually needs transcription, a reasoning model and a voice, plus a layer that handles interruptions and phone lines. Buying those from one vendor can cut weeks of integration work and leave fewer contracts to manage.
Larger companies often prefer to mix components. They may keep one vendor for transcription because of accuracy, another for voice because of brand fit, and a third for the reasoning model. A bundled stack is less attractive when each piece is already chosen.
The middle ground is to use the transcription models alone. Those have the strongest independent data, and they are easy to swap into an existing pipeline. If the results hold up on your audio, you can add the other pieces later.
Whichever route you choose, keep the audio pipeline modular. Models improve quickly, and a design that lets you replace one stage without rewriting the rest saves money over a year.
Should You Switch to Voxtral 3 for Voice Agents?
Maybe, but test first. The all-in agent price of $0.04 per minute is attractive compared with stitching together separate services. A single SDK can also cut integration work. Our guide on AI voice agents for production calls covers what teams should check before going live.
Watch for gaps. No time-to-first-audio is published for text-to-speech, and no maximum generation length is stated. Long-form coherence has not been tested publicly, so check it if you narrate long documents.
Follow this short list during a pilot.
- Run 100 real calls or recordings through each component.
- Measure word error rate on your own accents and jargon.
- Time a full agent turn, including your reasoning model.
- Confirm language, license and data handling terms in writing.
- Compute total cost per completed call, not per component.
One more practical point is audio quality. Models trained on clean studio recordings can struggle with phone-line audio, background noise and overlapping speakers. Always test with the messy audio you actually have, not with a demo clip. That test often matters more than a small difference in a published word error rate.
Verdict
Voxtral 3 is a strong, low-priced entry in speech, with independent accuracy data that supports the transcription claims for the Small and Mini models. The agent SDK and text-to-speech model are newer and less tested, and the language and license details need clarification. Model Drop's editors would pilot the transcription models now and wait for more arena results before betting a voice product on the synthesis model.
