Voxtral 3 Pricing and API: Transcription, TTS and Python Code

Updated 23 October 2026. Mistral released Voxtral 3, five speech models plus a voice agent SDK, on 20 October, and this guide covers prices, model IDs, working Python code and what breaks if you used Voxtral 2.

Voxtral 3 is Mistral's new family of speech models: three that turn speech into text and two that turn text into speech, all sold through the Mistral API. As of 23 October 2026, transcribing a recorded file costs $0.0035 per minute with Voxtral Mini Transcribe 3 and $0.004 per minute with Voxtral Small Transcribe 3, live transcription costs $0.008 per minute, and speech generation costs $13 per 1M characters.

That puts an hour of recorded audio at about 21 to 24 cents. The models label speakers, accept custom vocabulary and detect the language, but word-level timestamps are gone, which is the main breaking change for Voxtral 2 users. As of 23 October 2026, Mistral has not published Voxtral 3 weights on Hugging Face, so it is an API product unless you arrange a private deployment with Mistral.

Related: MiniMax M3.1 Flash: Price, Open Weights, API and vs DeepSeek

Voxtral 3 models, IDs and prices

From Mistral's announcement, its audio docs and the model cards, checked on 23 October 2026:

Use the full -3-0 IDs in production. Mistral's migration page also lists -latest aliases, which move when Mistral changes the current model; voxtral-mini-tts-latest, for example, will point to Voxtral 3 within a month of general availability.

How much does Voxtral 3 cost?

Mistral bills transcription per minute of audio and speech generation per character. As of 23 October 2026, per hour of audio that works out to:

• Voxtral Mini Transcribe 3: $0.21 per hour
• Voxtral Small Transcribe 3: $0.24 per hour
• Voxtral Mini Transcribe Realtime 3: $0.48 per hour
• Voice agent API: $2.40 per hour with Mistral Small 4, or $4.80 with a larger model

Example: a month of support calls

An example with assumed numbers: a support team records 2,000 hours of calls a month (120,000 minutes) and wants every call transcribed with speaker labels.

• Voxtral Mini Transcribe 3: 120,000 minutes x $0.0035 = $420 a month
• Voxtral Small Transcribe 3: 120,000 minutes x $0.004 = $480 a month

Summaries cost extra, because they run on a text model after transcription. For speech, assume a 1,000-word script is about 6,000 characters: 1,000 such scripts would cost about $78 with Voxtral TTS 3.

Voxtral 3 vs OpenAI and ElevenLabs prices

List prices from each vendor's pricing page as of 23 October 2026. OpenAI labels its per-minute figures as estimates, and ElevenLabs prices speech to text per hour:

The gap on file transcription is small, so choose on accuracy for your audio, language coverage and where the data is processed; ElevenLabs lists more than 90 languages for Scribe v2, against 50 for Voxtral. Text to speech is where Voxtral is clearly cheaper on list price. Rates come from the OpenAI pricing page and ElevenLabs API pricing.

How to use the Voxtral 3 API in Python

Create a key in Mistral's AI Studio, store it in the MISTRAL_API_KEY environment variable and install the SDK. Add the realtime extra only if you need live transcription over a WebSocket:

pip install mistralai
export MISTRAL_API_KEY="your-api-key"

Transcribe a file with speaker labels and custom vocabulary

This uses the parameters in Mistral's diarization guide and context biasing guide. Diarization needs segment timestamps, and context_bias takes names and terms that must be spelled your way. The file name and terms are examples:

import os
from mistralai.client import Mistral

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])

with open("sales_call.mp3", "rb") as audio:
    result = client.audio.transcriptions.complete(
        model="voxtral-mini-transcribe-3-0",
        file={"file_name": "sales_call.mp3", "content": audio},
        diarize=True,
        timestamp_granularities=["segment"],
        context_bias=["Acme Plumbing", "QuickBooks", "Dr. Okafor"],
    )

for seg in result.segments:
    print(round(seg.start, 1), seg.speaker_id, seg.text.strip())

Speaker labels come back as speaker_1, speaker_2 and so on, and only hold within one file. To get real names, send up to 8 reference clips of 3 to 15 seconds each. Pin the language with a two-letter code such as language="fr", or leave it out and the model detects it.

Generate speech with Voxtral TTS 3

The speech endpoint returns base64 audio. This version of Mistral's quickstart saves a WAV file instead of playing it, with a preset voice ID from that page:

import base64
import os
from mistralai.client import Mistral

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])

response = client.audio.speech.complete(
    model="voxtral-mini-tts-3-0",
    input="Thanks for calling. Your order shipped this morning.",
    voice_id="gb_jane_neutral",
    response_format="wav",
)

with open("reply.wav", "wb") as f:
    f.write(base64.b64decode(response.audio_data))

Generated speech is mono at 24 kHz, and response_format accepts mp3 (the default), opus, wav, flac or pcm. Voxtral TTS 3 adds 34 inline emotion tags, slow and fast pace tags, phoneme hints for names, and voice cloning from an MP3 or WAV clip of a few seconds to a few minutes.

What changed from Voxtral 2, and what breaks

The migration page maps voxtral-mini-2602 to Voxtral Mini Transcribe 3, voxtral-small-2507 to Voxtral Small Transcribe 3, voxtral-mini-transcribe-realtime-2602 to the realtime 3 model and voxtral-mini-tts-2603 to Voxtral TTS 3. The changes that matter:

One more gap: Voxtral Small Transcribe 3 only transcribes at launch. To summarize or translate a recording, Mistral's docs point to audio in chat completions with voxtral-small-latest, a Voxtral 2 model.

Is Voxtral 3 better than Whisper?

On Mistral's own charts it is ahead, but these are vendor results. Mistral says Voxtral Small Transcribe 3 outperforms Whisper, ElevenLabs Scribe v2 and Gemini 3.5 Transcribe on multilingual transcription. Its diarization chart, measured as cpWER across five public benchmarks (lower is better), shows:

On the Artificial Analysis word error rate index that Mistral plots on the same page, Voxtral Small Transcribe 3 scores 2.10, behind Microsoft AI Transcribe 2 (2.04) and Fun-Realtime-ASR-preview (1.73), and ahead of Scribe v2 (2.18) and Whisper large-v3 (4.07). For live audio, Mistral's FLEURS chart shows the realtime model's error rate rising from 5.7 at a 2,400 ms delay to 7.4 at 240 ms. Test on 20 to 50 of your own recordings before you switch.

Which languages does Voxtral 3 support?

As of 23 October 2026, Mistral's languages page lists 50 languages for file transcription with both the Mini and Small models, which the announcement says include the 24 official EU languages, and 45 languages and dialects for realtime transcription. Voxtral TTS 3 supports 35 languages with more than 50 regional dialects, with preset voices for 22 language and accent options and voice cloning for the rest. The announcement quotes lower counts for realtime languages and presets, so check the languages page for yours.

Voice agents: the third part of Voxtral 3

The voice agent SDK, built on LiveKit, chains realtime transcription, a reasoning model such as Mistral Small and Voxtral TTS into one streaming pipeline. Mistral says each turn completes in under one second and that the agent can call functions and query your own systems, over SIP phone lines or WebRTC in browsers and apps. As of 23 October 2026 it costs $0.04 per minute with Mistral Small 4 and $0.08 with a larger model. For the business side of a phone assistant, see our AI voice receptionist page.

Can you self-host Voxtral 3?

Not from a public download. As of 23 October 2026, a Hugging Face search shows only older Mistral Voxtral checkpoints, such as Voxtral-Mini-4B-Realtime-2602 and Voxtral-4B-TTS-2603, and the announcement asks companies that want the full stack on-premises or in a private cloud to contact Mistral. The API runs on European, US and global endpoints. Mistral's legacy docs list the older Voxtral 2 realtime model under Apache 2.0 and Voxtral TTS 2 under CC BY-NC 4.0, a non-commercial license. Our comparison of open-weight models and API models covers the trade-offs, and our Mistral Large 4 guide covers Mistral's main text model.

Where Voxtral 3 fits in a business tool

• Meeting notes and call summaries. Transcribe with speaker labels, then summarize with a text model into a fixed template. Our AI call summary templates give six formats to start from, and an AI meeting notes summarizer can file the result in your CRM.
• Support quality checks. Reference clips label "agent" and "customer" in every call, so you can search what customers said across thousands of calls, in any of the 50 transcription languages.
• Spoken replies. At $13 per 1M characters, reading out order updates, alerts or help articles is cheap enough to test widely.

Treat transcripts like the recordings they come from: follow the same consent rules, limit who can read them and set a retention period. To budget the text-model side, see our guide to estimating LLM API costs. If you want transcription or summaries added to your own systems, send us the details.

Voxtral 3 FAQ

How much does Voxtral 3 cost?

As of 23 October 2026, Voxtral Mini Transcribe 3 costs $0.0035 per minute of audio, Voxtral Small Transcribe 3 costs $0.004 per minute and Voxtral Mini Transcribe Realtime 3 costs $0.008 per minute. Voxtral TTS 3 costs $13 per 1M characters. The voice agent API costs $0.04 per minute with Mistral Small 4 or $0.08 per minute with a larger model, all billed through the Mistral API.

What is the Voxtral 3 model name in the API?

Use voxtral-mini-transcribe-3-0 or voxtral-small-transcribe-3-0 for recorded files, voxtral-mini-transcribe-realtime-3-0 for live audio and voxtral-mini-tts-3-0 for speech generation. Each also answers to a shorter alias, such as voxtral-mini-transcribe-3. Mistral's migration page shows which Voxtral 2 ID each one replaces, such as voxtral-mini-2602, so update your code and test again.

Is Voxtral 3 open source?

Not at launch. As of 23 October 2026, Mistral offers Voxtral 3 through its API on European, US and global endpoints, and Hugging Face lists only older Voxtral models from Mistral. The announcement says companies can run the full stack on-premises or in a private cloud by contacting Mistral. Some earlier Voxtral models, such as the 2602 realtime model, were released with open weights under Apache 2.0.

Does Voxtral 3 give word-level timestamps?

No. Voxtral 3 file transcription returns timestamps per segment only, and a request for word timestamps returns a 400 error. This is the main breaking change from Voxtral 2, which offered word-level timestamps. Realtime segments carry a speaker label but no start or end time. If your product highlights words as audio plays, plan a replacement for that feature before you migrate.

How many languages does Voxtral 3 support?

Mistral's languages page lists 50 languages for file transcription, 45 languages and dialects for realtime transcription, and 35 languages with more than 50 regional dialects for text to speech. You can pin a language with a two-letter ISO code such as fr, or leave it out and let the model detect it. Preset voices cover fewer languages, so use voice cloning for the rest.

Is Voxtral 3 better than Whisper?

Mistral says Voxtral Small Transcribe 3 beats Whisper, ElevenLabs Scribe v2 and Gemini 3.5 Transcribe on multilingual transcription, and its diarization chart shows lower error than Scribe v2 and Gemini 3.5 Transcribe. On the Artificial Analysis index that Mistral plots, Voxtral Small 3 scores 2.10 against 4.07 for Whisper large-v3. These are vendor-selected results, so compare both on your own recordings first.

Sources

• Mistral AI: Voxtral 3 announcement
• Mistral docs: Audio overview
• Mistral docs: Speech to Text
• Mistral docs: Diarize speakers
• Mistral docs: Bias context and vocabulary
• Mistral docs: Text to Speech
• Mistral docs: Quickstart, generate speech
• Mistral docs: Migrate from Voxtral 2
• Mistral docs: Languages
• Mistral docs: Limits
• Mistral docs: Audio formats
• Mistral docs: Audio (legacy Voxtral 2 overview)
• Mistral model card: Voxtral Mini Transcribe 3
• Mistral model card: Voxtral Small Transcribe 3
• Mistral model card: Voxtral Mini Transcribe Realtime 3
• Mistral model card: Voxtral TTS 3
• OpenAI API pricing
• ElevenLabs API pricing
• Hugging Face: Mistral Voxtral models
• Hacker News: Voxtral 3 launch discussion (20 October 2026)