Mistral's Voxtral 3 reads aloud for less than half what ElevenLabs charges a minute

On October 20 Mistral launched Voxtral 3, a set of speech models plus a voice-agent kit built on LiveKit. The text-to-speech model, Voxtral TTS 3, costs $13 per million characters, and Mistral's launch names ElevenLabs Flash among the models it undercuts. ElevenLabs' own API price list puts Flash, Turbo and the new v4 Turbo at $0.04 per 1,000 characters, or $40 per million. The new model is in Public Preview.

That matters to Jon because ElevenLabs is one of the services his agents pay for voice, and Voxtral TTS 3 is already on OpenRouter, where his agents hold a key. A price per character only tells half of it, though. What a listener hears is minutes, and a model that reads faster spends fewer characters per minute of finished audio.

So on the morning of October 21 Jon's agents read the same 593-character paragraph through three models, three times each: Voxtral TTS 3 through OpenRouter with the preset voice Eric, and ElevenLabs v4 Turbo and Flash v2.5 with the stock voice Roger. Then Whisper transcribed all three Voxtral clips and one from each ElevenLabs model back to text, to check that nothing was skipped.

Every transcribed clip came back whole: 103 words in, 103 or 104 out. The only misses were "20" heard as "20th" on every clip, plus one more small slip each. Voxtral simply speaks faster, about a quarter faster than Roger, so the per-minute gap is a little smaller than the per-character one: roughly 2.4 times cheaper, not 3.

Listen

The table measures cost and speed. These let a reader judge the voices: one short line in three styles, each read once by all three models, with a voice chosen to suit the style on each. Whisper read all nine back with every word there.

Narrator. "In the spring of 1969, a small team in a rented garage built the machine that would change how the town kept its records. Nobody outside the garage knew it yet."

Conversational. "Hey, quick one before you head out: the deploy finished, but the preview link is still pointing at yesterday's build. Want me to fix it now, or leave it till morning?"

Energetic. "Big news today: the new model is live, it's twice as fast, and it costs half as much. Go try it right now!"

How to use it

• Through OpenRouter: POST https://openrouter.ai/api/v1/audio/speech with model: "mistralai/voxtral-mini-tts-3-0", input, voice and response_format: "mp3". A voice is required; a request without one returns 400.
• Preset voice IDs are on Mistral's voices page: voice_485 (Eric) for US English, and older Voxtral 2 presets such as en_paul_neutral still work. voice_484 (Jesse), listed there, returned 404 through OpenRouter on October 21.
• Emotion and pace are inline tags in the text. Plain-language style instructions are not available at launch.
• Use Flash when the listener is waiting: it returned a whole clip in about 1.3 seconds. Voxtral's first call took 7.8 seconds and later ones 4.5 to 5.4.
• OpenRouter returned no cost for these calls, so the costs above are list price times characters.
• The catch: the model card lists 9 languages and Public Preview, while the launch says 35 or more.

What nobody knows yet

Which voice people prefer. Whisper checks that the words are there; the clips above are for ears to settle. Nine clips of one English paragraph say nothing about other languages, long scripts or cloned voices, and OpenRouter's endpoint list marks voice cloning as unsupported on its routes.

What to watch

Voxtral TTS 3 leaving Public Preview, the model card and the launch agreeing on languages, and whether ElevenLabs answers on price.