NewModelHubBuild Your Own Stack — Component Cost
Cached input is billed at one tenth of the standard rate on Mistral's pricing table (e.g. $0.0004 instead of $0.004 per minute). Leave the cached share at 0 if you are not using cached prompts or context.
Estimated stack cost per month: $0.00
Voice Agent API — All-In Cost
Mistral's Voice Agent API bundles realtime transcription, reasoning with tool calling, and speech generation into a single per-minute price. No separate TTS or transcription line items — but the reasoning model choice doubles the rate.
Estimated Agent API cost per month: $0.00
Voxtral 3 Rate Card
| Model | Billed per | Standard | Cached input |
|---|---|---|---|
| Voxtral Mini Transcribe 3 (batch) | minute of audio | $0.0035 | $0.00035 |
| Voxtral Small Transcribe 3 | minute of audio | $0.004 | $0.0004 |
| Voxtral Mini Transcribe Realtime 3 | minute of audio | $0.008 | $0.0008 |
| Voxtral Speech 3 (TTS) | 1M characters | $13 | $1.30 |
| Voice Agent API · Mistral Small 4 | conversation minute | $0.04 | — |
| Voice Agent API · larger models | conversation minute | $0.08 | — |
| Voxtral Mini Transcribe 2 (previous gen) | minute of audio | $0.003 | $0.0003 |
| Voxtral TTS (previous gen) | 1M characters | $16 | $1.60 |
Rates from Mistral's official pricing page (docs.mistral.ai/inference/pricing) and the Voxtral 3 launch post (mistral.ai/news/voxtral-3), verified October 22, 2026. Mistral's pricing table lists the TTS model as "Voxtral TTS 3"; the launch post calls it Voxtral Speech 3 — same $13 per million characters. Voxtral 3 is served from European, US and global endpoints; on-premises and private-cloud deployment is via Mistral's sales team.
What Drives Your Voxtral 3 Bill
- Three units, one invoice. Minutes for transcription, characters for TTS, conversation minutes for the agent API. The classic budgeting error is pricing a voice agent as if only one unit applied — a one-hour agent call is 60 transcription minutes plus roughly 24,000–48,000 TTS characters plus the reasoning model.
- Batch vs realtime is a 2.3× decision. Mini Transcribe 3 batch costs $0.0035/min; the realtime model costs $0.008/min. Anything that is not a live conversation — meetings, archives, podcasts — belongs on the batch or Small tier.
- Characters are not minutes. At an average speaking rate of ~800 characters per minute, one million TTS characters is about 21 hours of speech — $13, or roughly $0.62 per hour of generated audio. Verbose agent scripts inflate this fast: a reply that talks twice as long costs twice as much.
- The Agent API bundles the LLM — DIY does not. Speech components alone run about $0.0132 per conversation minute, well under the $0.04 Agent API rate. The gap is the reasoning model, tool calling and pipeline operations. If your LLM spend per minute is small, DIY wins; if you would otherwise run a frontier model per turn, the bundle can be cheaper.
- Cached input is a 90% discount. Repeated context — fixed prompts, glossaries, recurring speakers' context — billed as cached input costs one tenth of standard on both transcription and TTS. On a 100,000-minute Small workload with 30% cached, that is $292 instead of $400.
- Accuracy is nearly free here. Small Transcribe 3 costs just $0.0005/min more than Mini batch — $30 per 100,000 minutes. Unless your volume is enormous, buy the accuracy tier and spend your optimization effort on the realtime-vs-batch split instead.
FAQ
How much does Voxtral 3 cost?
Transcription is billed per minute of audio: Mini Transcribe 3 (batch) $0.0035/min, Small Transcribe 3 $0.004/min, Mini Transcribe Realtime 3 $0.008/min. Text-to-speech (Voxtral Speech 3) is $13 per million characters. The all-in Voice Agent API is $0.04 per conversation minute with Mistral Small 4, or $0.08/min with larger models.
What is the difference between the three transcription models?
Mini Transcribe 3 is the high-volume batch option at $0.0035/min. Small Transcribe 3 ($0.004/min) is the accuracy tier — 4.0% word error rate on Mistral's FLEURS test. Mini Transcribe Realtime 3 ($0.008/min) is for live agents, with sub-500ms streaming latency and end-of-turn detection.
What does the Voice Agent API include?
The full pipeline — realtime transcription, a reasoning LLM with tool calling, and speech generation — at one per-minute price ($0.04 with Mistral Small 4, $0.08 with larger models). Assembling the components yourself, the speech parts cost about $0.0132 per conversation minute, but the reasoning LLM is billed separately on top.
What is cached input pricing on Voxtral 3?
Mistral's pricing table lists cached input at one tenth of the standard rate for the speech models — $0.0004 instead of $0.004 per minute for Small Transcribe 3, $1.30 instead of $13 per million TTS characters. The cached rate applies to the cached portion of your input.
Is Voxtral 3 cheaper than the previous Voxtral generation?
On TTS, yes: $13 per million characters versus $16 for the previous Voxtral TTS. Entry transcription moved slightly up — Mini Transcribe 2 was $0.003/min versus $0.0035 for Mini Transcribe 3 batch — in exchange for the newer models' accuracy and features.
How accurate is Voxtral Small Transcribe 3?
Mistral reports a 4.0% word error rate on its FLEURS test across 45 languages (language inferred), versus 14.9% for Whisper large-v3 on the same test. These are Mistral's published figures on clean, read speech — budget for higher error rates on noisy real-world audio such as phone calls.
Methodology & sources: all rates verified against Mistral's official pricing page and the Voxtral 3 launch post on October 22, 2026. The DIY-vs-Agent comparison assumes the agent speaks about half of each conversation minute at ~800 characters per minute of speech (≈400 TTS characters per conversation minute) and excludes your LLM cost, which the component stack bills separately. Estimates exclude taxes; your invoice follows metered usage.