Voxtral 3 TTS Pricing: Check the Model Before You Budget
Checked 20 October 2026. Mistral’s Voxtral 3 launch page lists speech generation at $13 per million characters, while its linked Voxtral TTS 3 model card displays $16 per million characters. That is a documentation discrepancy—not proof that either rate is the amount your account will be billed. Before quoting a customer or approving a rollout, confirm the exact model, provider and applicable rate.
What the primary pages actually show
These are observations from public pages, not an account invoice or a tested API charge. The two model names also matter: the announcement calls the speech product Voxtral Speech 3; the linked model card is titled Voxtral TTS 3 and identifies voxtral-mini-tts-3-0.
| Source checked | Product label | Displayed TTS rate | Important limit |
| --- | --- | --- | --- |
| [Voxtral 3 announcement](https://mistral.ai/news/voxtral-3/) | Voxtral Speech 3 | $13 / 1M characters | Launch-page figure; not a verified account quote |
| [Voxtral TTS 3 model card](https://docs.mistral.ai/models/voxtral-mini-tts-3-0) | voxtral-mini-tts-3-0 | $16 / 1M characters | Marked Public Preview / Premier v3.0 |
| [General API pricing](https://docs.mistral.ai/inference/pricing) | Voxtral TTS | $16 / 1M characters | Generic family label; do not assume it resolves the v3 question |
We could not establish why the pages differ or independently verify the effective billing rate. Do not describe this as a confirmed price cut, a billing error, or a guaranteed discount. Recheck the dated pages and ask the service provider to confirm which commercial terms apply to the model you will call.
Turn a headline price into a usable estimate
Create a small cost sheet before choosing a rate. Record the provider you pay, the exact model identifier, the endpoint or deployment route, the region and service tier where relevant, the charging unit, and the date of the quote. Keep a copy of the applicable pricing or written confirmation alongside the estimate. An alias containing latest is less useful for a reproducible budget than a recorded model/version and a known alias policy.
Then estimate volume using your own scripts. Count the text your application actually submits for speech generation, including repeated answers or regenerated clips. Do not estimate the entire project from the number of calls alone: a short confirmation and a long explanation can have very different text lengths. Confirm the provider’s character-counting and retry rules rather than assuming how punctuation, failed requests or cached content are treated.
A transparent two-rate scenario—not a bill
For an illustrative workload of two million billable characters, a $13-per-million rate produces $26 of TTS usage; a $16-per-million rate produces $32. The $6 difference is simple arithmetic showing why the source mismatch matters. It is not a quote, a claim about your character count, or an estimate of the complete voice-agent project. Taxes, discounts and other services are not included.
TTS is only one part of a voice agent
Mistral describes a pipeline with speech recognition, a separate reasoning model and speech output. A character-based TTS line therefore should not be presented as the whole conversation cost. Budget separately for transcription, reasoning, telephony or transport, connected services and your own operating work where applicable. The launch directs teams seeking the full on-premises or private-cloud stack to contact Mistral; that is not an automatic public-weight licence.
A practical pilot handoff
Use a small, authorised test set before scaling: one short response, one longer response, a repeated request and an interrupted conversation. Record the model selected, the usage reported and the charges when they become available. Reconcile those records with the provider’s terms before multiplying the pilot result into a monthly forecast. We have not performed that billing test for this article.
For a customer proposal, label unconfirmed prices as assumptions and show which costs are excluded. Keep the first pilot reversible: a low spend limit, a narrow use case and a human review point make a mistaken assumption easier to detect. The useful decision today is not to guess which public page wins; it is to make the estimate traceable to the service you will actually buy.
Two checks before approving a voice budget
Sources and scope
Primary sources: [Mistral’s Voxtral 3 launch](https://mistral.ai/news/voxtral-3/), [Voxtral TTS 3 model card](https://docs.mistral.ai/models/voxtral-mini-tts-3-0), [API pricing](https://docs.mistral.ai/inference/pricing), and [Voxtral Mini Transcribe Realtime 3](https://docs.mistral.ai/models/voxtral-mini-transcribe-realtime-3-0). Public-page comparison only; no hands-on quality benchmark or verified account billing result is claimed. Prices and preview status can change.
