Mistral’s news page has no clock. It only says the models are available today. The chart notes say the figures were generated on 2026-10-20. The judgment first: the prices are in the table, and the word-error rates are not in the prose.

The transcription model is named Voxtral Small Transcribe 3. The page says the language list now has 50, including 24 official EU languages and many low-resource Asian and Indic languages. It also says that on English, French, and Spanish transcription and on diarization, Small outperforms Whisper, Scribe v2, and Gemini 3.5 Transcribe. The comparisons are charts. The notes name a macro-average word-error rate over 45 FLEURS languages, and AMI-IHM, AMI-SDM, Switchboard, Switchboard CallHome, and AISHELL-4. The prose prints none of the scores.

[1]
A closed telephone handset on its base
A closed telephone handset resting on its base, charcoal body, muted blue edge., AI-generated illustration, not a news photograph

Mini has 4B parameters and costs $0.0035 a minute. In the price table, batch Mini Transcribe 3 is $0.0035 a minute and Small Transcribe 3 is $0.0040 a minute.

The prose names the realtime model Voxtral Realtime Transcribe 3 and says it runs in over 35 languages at a latency under 500 milliseconds. The card and the price table name it Voxtral Mini Transcribe Realtime 3, at $0.008 a minute, and put the language count at 40 or more. The page does not say whether the two names are the same model. Both sentences are the company’s.

[1]

Speech synthesis has two names as well. The highlights say Voxtral Mini TTS 3. A later section says Voxtral Mini TTS 3.0 and puts the price on Voxtral Speech 3: $13 per million characters. The price table uses Speech 3 too, with batch and realtime on the same API. The page says it supports 35 or more languages and 50 or more dialects, and that this price is lower than OpenAI TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash. Those four prices are not printed here.

Voice Code Bench is BeSimple AI’s, aimed at numbers and unusual names in enterprise support. The note says 300 English prompts, two voices per system, judge gemini-3.8-flash, prices from vendor pages, generated 2026-10-20. The scores stay in the chart. Other charts labeled as internal evaluations also leave their scores out of the prose.

There are two sentences about duration. Do not merge them. Pro Voice says 30 minutes of audio is enough to reproduce a speaker’s character, emotional range, and speaking style across languages. Another line says the features include voice matching from 30 or more minutes of sampled audio. Both wordings are the page’s. This piece does not describe how a voice is made from that audio.

[1]

The voice-agent SDK is built on LiveKit. The page says listening, reasoning, and speaking are one stream, over the phone or the web. The listening example is Mini realtime transcription. The reasoning example is Mistral Small. The reply is Speech 3. A separate price row is Mistral Small 4 at $0.04 a minute, and any other reasoning model at $0.08 a minute, printed as 0.04$/min and 0.08$/min. Small and Small 4 both appear. The page does not say whether they are the same model.

The page says each turn completes in under one second. It does not give measurement conditions. The call examples are verifying identity, checking a bill, and booking an appointment. That is a list, not a test in this piece.

Availability is two sentences. Do not merge them. The highlights say it is available today through the EU-hosted API or on your own infrastructure. The closing section says the hosted endpoints include Europe, the US, and global. On-premises or private cloud means talking to their team, with no public price.

[1]