Model Releases · OCTOBER 21, 2026
Voxtral 3 lands at 2.1% WER on Artificial Analysis, prices TTS at $13/1M chars, and bundles a LiveKit voice-agent SDK
Mistral's October 20 release puts a 4B transcriber third on the Artificial Analysis speech-to-text board, undercuts named TTS incumbents at $13 per million characters, and prices a full agent loop at $0.04 per minute with Mistral Small.
Mistral shipped the Voxtral 3 family on October 20, and within a day the Artificial Analysis speech-to-text board was reading Voxtral Small Transcribe 3 at 2.1% WER, third place behind Alibaba's Fun-Realtime-ASR-preview at 1.7% and Microsoft's MAI-Transcribe-2 at 2.0%. ElevenLabs Scribe v2 trails by a tenth of a point at 2.2%. The release is three products at once: a transcription stack that's suddenly competitive at the frontier, a TTS model priced to pressure the field, and a LiveKit-powered agent SDK that collapses what used to be a three-vendor integration into one.
The transcription number is the one that matters. Voxtral Mini Transcribe 3, the 4B-parameter sibling, also scores 2.2% WER and sells for $0.0035 per minute in batch, 12.5% below the Small tier's $0.0040. The realtime variant lists at $0.008 with claimed sub-500ms latency. All transcription models cover 50+ languages, including the 24 official EU languages, which is the part of the pitch aimed squarely at Mistral's home-field regulators.
Speed is the asterisk. OrcaRouter clocks Small Transcribe 3 at a median speed factor of 66.6, roughly 67 minutes of audio processed per minute of compute. That's about four times slower than MAI-Transcribe-2 at 308.8, though three times faster than Fun-Realtime-ASR-preview at 20.3. Mistral's own evals report 81.4% keyphrase recall on Earnings-22 without context biasing and 94.6% with it, plus 18.4% cpWER on diarization averaged across AMI-IHM, AMI-SDM, Switchboard, CallHome, and AISHELL-4.
The TTS side is a different conversation. Voxtral Speech 3, also marketed as Voxtral Mini TTS 3.0, lists at $13 per million characters and sits at #13 on the Provider Voice arena with 1,217 Elo (95% CI 1,202–1,232, 1,619 comparisons), tied-band 11 through 17. On the Controlled Voice arena it's #14 at 1,066 Elo. ElevenLabs Eleven v4 is at 1,328, Qwen-Audio-3.1-TTS-Plus at 1,299, Cartesia Sonic 3.6 at 1,289. Mid-pack quality, but the 127-point jump from March 2026's Voxtral TTS at 1,090 is the trajectory worth tracking.
The SDK is the structural piece. Bundling transcription, a reasoning layer (Mistral Small 4, following the Large 4 preview earlier this month), TTS, and LiveKit transport into one agent loop priced around $0.04 per minute removes the glue work that kept voice agents in the hands of teams who could afford to stitch three vendors together. That's the move. The leaderboard position is marketing; the integration surface is the product.
Sources
- https://mistral.ai/news/voxtral-3/
- https://www.testingcatalog.com/mistral-launches-voxtral-3-with-a-voice-agent-sdk/
- https://www.orcarouter.ai/blog/voxtral-3-speech-family-release
- https://explainx.ai/blog/voxtral-3-mistral-voice-agent-sdk-speech-models
- https://fourweekmba.com/ai-mistrals-voxtral-3-transcribes-speech-from-0-0035-a-minute/