# Mistral’s Voxtral 3 Debuts at $0.0035 a Minute [2026] > Source: > Published: 2026-10-22 05:15:51+00:00 Mistral AI pushed a new product line into the open on October 20, 2026, and it has nothing to do with chatbots. The Paris-based lab launched **Voxtral 3**, a family of speech-to-text and text-to-speech models paired with a real-time voice agent SDK, and priced the whole stack down to fractions of a cent per minute. The launch, detailed on [Mistral’s official announcement page](https://mistral.ai/news/voxtral-3/), lands at a moment when voice has quietly become one of the most contested fronts in AI product design, sitting behind call centers, in-car assistants, and the next wave of phone-based customer service bots. What makes Voxtral 3 worth tracking is not a single headline number. It is the combination of aggressive per-minute pricing, a LiveKit-based SDK built for live conversation rather than batch processing, and a set of claims that put Mistral directly up against OpenAI, Google, ElevenLabs, Deepgram, and Microsoft in the same breath. Mistral is betting that voice AI buyers care more about cost per minute and latency than brand name, and Voxtral 3 is the clearest test of that bet yet. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. [Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F) ## What Mistral Actually Shipped on October 20 Voxtral 3 is not one model. It is a family of four named products, each aimed at a different slice of the speech pipeline. Mistral’s own description, confirmed on its governance and documentation pages, frames the launch as a way to connect listening, reasoning, tool use, and speaking into a single streaming system rather than four separate services bolted together after the fact. That framing matters because most voice products on the market today still stitch together a transcription API, a separate language model, and a third text-to-speech vendor, with latency stacking up at every handoff. The four confirmed products are Voxtral Mini Transcribe 3 for batch speech-to-text, Voxtral Mini Transcribe Realtime 3 for streaming speech-to-text, Voxtral Small Transcribe 3 as the flagship transcription model with added translation and summarization features, and Voxtral Mini TTS 3 for speech generation. Mistral’s legal and governance page lists all four as active, specialized models dated to the October 20 release, which is the clearest official confirmation that this is a coordinated product launch rather than a single experimental drop. One detail Mistral has not clarified is whether “Voxtral 3” refers to a single underlying model or simply the brand name for the whole product family. Available documentation treats it as the latter: a platform name covering four distinct, separately priced products rather than one model with four modes. That distinction matters for developers trying to understand what they are actually buying before they integrate. ## Meet the Voxtral 3 Model Family Voxtral Mini Transcribe 3 is the smallest and most concrete of the four, confirmed to run on 4 billion parameters and aimed at batch transcription jobs where a few seconds of extra latency does not matter. Voxtral Mini Transcribe Realtime 3 is the streaming cousin, built for live captioning and voice agents, with configurable latency that Mistral says can run anywhere from 80 milliseconds up to 2,400 milliseconds depending on how a developer tunes the tradeoff between speed and accuracy. It also adds end-of-turn detection and speaker diarization, both of which are the kind of unglamorous features that decide whether a voice agent feels natural or feels like talking to a bad phone tree. Voxtral Small Transcribe 3 sits above both Mini models and is what Mistral calls its flagship transcription product. Beyond raw speech-to-text, it adds translation, speech-based question answering, summarization, and function calling, which pushes it closer to being a reasoning layer that happens to accept audio input rather than a pure transcription tool. Voxtral Mini TTS 3 closes the loop on the speaking side, with zero-shot and professional voice cloning, a roughly 90-millisecond time-to-first-audio figure, five output audio formats, and a notably specific set of 34 emotion tags plus two pacing tags for controlling how a synthesized voice delivers a line. Mistral’s marketing material describes Voxtral 3 as supporting real-time conversation across more than 50 languages, while a separate figure for Voxtral Mini TTS 3 cites 35-plus languages and 50-plus dialects for expressive speech generation, with a narrower nine-language figure appearing elsewhere in the documentation. The gap between those numbers is not resolved in Mistral’s public materials, and developers evaluating the platform for a specific market should confirm exact language support for their use case directly with Mistral rather than relying on the headline figure. ## Inside the Pricing: What a Minute of AI Speech Costs Now Pricing is where Voxtral 3 is making the loudest noise. Mistral published per-minute and per-character rates for all four products plus the combined voice agent API, and the numbers undercut most publicly listed competitor pricing by a wide margin on the transcription side. | Product | Function | Price | Key confirmed specs | |---|---|---|---| | Voxtral Mini Transcribe 3 | Batch speech-to-text | $0.0035 per minute | 4 billion parameters, multilingual | | Voxtral Mini Transcribe Realtime 3 | Streaming speech-to-text | $0.008 per minute | 80ms-2,400ms configurable latency, diarization | | Voxtral Small Transcribe 3 | Flagship transcription | $0.0040 per minute | Translation, speech Q&A, summarization, function calling | | Voxtral Mini TTS 3 | Text-to-speech | $13 per 1 million characters | ~90ms time-to-first-audio, 34 emotion tags, voice cloning | | Voice Agent API (Mistral Small 4 reasoning) | Full voice agent | $0.04 per minute | LiveKit-based streaming pipeline | | Voice Agent API (larger reasoning model) | Full voice agent | $0.08 per minute | LiveKit-based streaming pipeline | Run the math on a typical customer service call and the appeal becomes obvious. A five-minute support call handled end to end through the cheaper voice agent tier costs roughly 20 cents in model fees, excluding telephony and infrastructure costs. That is the kind of per-interaction economics that makes a product team willing to rip out an existing vendor relationship, assuming the quality holds up under real traffic rather than demo conditions. ``` # Rough cost estimate for a 5-minute support call minutes=5 rate_per_minute=0.04 # Voice Agent API, Mistral Small 4 reasoning tier echo "scale=2; $minutes * $rate_per_minute" | bc # Output: .20 (approx $0.20 per call, before telephony costs) ``` Mistral has not published separate per-token pricing for any Voxtral 3 product, which keeps the billing model simple for buyers but also makes it harder to compare directly against vendors who price by token rather than by minute or by character. For a marketing or support team modeling a new voice feature, this is the one gap in the published materials worth flagging to a finance partner before signing a contract. ## The Voice Agent SDK and Why LiveKit Matters The SDK is arguably the more consequential half of the launch. Mistral AI, in its official announcement, put it this way: “Voxtral 3 also introduces a voice agent SDK, built on LiveKit. Transcribe listens, a reasoning model thinks and calls tools, and Speech replies, streamed as one pipeline for natural, intelligent conversation.” That is a direct architectural statement, not marketing fluff, and it tells developers exactly what they are integrating: a transcription layer, a reasoning layer, and a speech layer, all running through a single streaming connection rather than three separate API calls stitched together by application code. LiveKit is a widely used open infrastructure layer for real-time audio and video, and building the SDK on top of it rather than a proprietary transport signals that Mistral wants developers who already know LiveKit’s tooling to be able to drop Voxtral 3 in without relearning an entire networking stack. In a separate part of its launch materials, Mistral described the same idea slightly differently: “To address this, we introduce Voxtral 3: a new family of state-of-the-art speech-to-text and text-to-speech models, alongside a voice agent SDK that brings listening, thinking, tool calling, and speaking into a single real-time streaming pipeline.” The repetition of “single pipeline” across both statements is the clearest signal of what Mistral thinks its actual competitive edge is: not raw transcription accuracy, but end-to-end latency. What the public materials do not confirm is just as important for engineering teams. There is no published SDK package name, no confirmed list of supported programming languages, no version number, no public repository location, and no stated service-level guarantee for production use. Mistral also has not said whether LiveKit is a mandatory dependency or an optional transport layer that can be swapped out. Teams evaluating Voxtral 3 for anything beyond a prototype should expect to request this information directly rather than finding it in public docs today. ## How Voxtral 3 Stacks Up Against OpenAI, ElevenLabs, and Google Mistral is not shy about naming names. The company claims Voxtral Small Transcribe 3 outperforms Whisper large-v3, GPT-Transcribe, Scribe v2, Gemini 3.5 Transcribe, and Microsoft AI Transcribe 2. On the speech generation side, Mistral characterizes Voxtral Mini TTS 3 as cheaper than comparable offerings from OpenAI TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash. Those are five and four named competitors respectively, which is an unusually direct set of comparisons for a company to put in a launch post. | Competing product | Type | Mistral’s stated comparison claim | |---|---|---| | OpenAI Whisper large-v3 | Open-weight transcription baseline | Claims Voxtral Small Transcribe 3 outperforms it | | OpenAI GPT-Transcribe | Closed, proprietary transcription | Claims Voxtral Small Transcribe 3 outperforms it | | Google Gemini 3.5 Transcribe | Closed, multimodal speech | Claims Voxtral Small Transcribe 3 outperforms it | | Microsoft AI Transcribe 2 | Closed, enterprise transcription | Claims Voxtral Small Transcribe 3 outperforms it | | ElevenLabs Flash | Closed, low-latency TTS | Claims Voxtral Mini TTS 3 is cheaper | | OpenAI TTS-1 / Deepgram / Gemini Flash | Closed TTS and speech APIs | Claims Voxtral Mini TTS 3 is cheaper on all three | None of these comparisons come with published benchmark scores, error rates, test datasets, or confidence intervals in the materials reviewed for this article. That does not make the claims false, but it does mean every number in that table above is Mistral’s framing of itself against rivals rather than an independently verified result. Buyers evaluating Voxtral 3 against an incumbent vendor should run their own side-by-side test on their actual audio data before trusting a launch post’s comparison table, a caveat that applies equally to every AI vendor’s launch marketing, not just Mistral’s. ### Where the Comparisons Hold Up and Where They Don’t The pricing comparisons are easier to check than the quality comparisons because pricing is a published number on both sides, while transcription accuracy depends heavily on audio quality, accent, background noise, and domain-specific vocabulary. A support line fielding calls from a single English-speaking market will see very different results than a global call center handling a dozen languages and heavy accents, and Mistral’s launch materials do not break out performance by language or acoustic condition. ## Latency Numbers That Matter for Real-Time Voice Latency is the metric that actually decides whether a voice agent feels usable on a phone call, and here Mistral has published real figures rather than vague claims. Voxtral Mini Transcribe Realtime 3 offers configurable latency from 80 milliseconds up to 2,400 milliseconds, letting a developer trade accuracy for speed depending on the use case. A live interpreter application might tolerate the lower end of that range poorly and need the full 2.4 seconds for accuracy, while a simple voice command system might run fine at 80 milliseconds. On the output side, Voxtral Mini TTS 3’s roughly 90-millisecond time-to-first-audio figure is the number that determines how quickly a caller hears a response start after an agent decides what to say. Humans notice gaps above roughly 200 to 300 milliseconds in live conversation as unnatural, so a sub-100-millisecond figure, if it holds up under real network conditions rather than lab benchmarks, puts Voxtral 3 in a competitive position against the fastest commercial TTS systems on the market today. ## What Mistral Is Not Confirming Yet Several details that typically matter most to engineering and legal teams remain unconfirmed in Mistral’s public materials as of this writing. The licensing status of any Voxtral 3 model, whether open-weight or fully proprietary, is not stated in the governance page reviewed for this piece. Context window size for any of the four models is not published. Parameter counts for three of the four models, everything except the 4 billion figure for Voxtral Mini Transcribe 3, are not disclosed. Whether the API has distinct EU, US, and global endpoints, as a third-party report has claimed, is not corroborated in Mistral’s own official pages. None of this is unusual for a same-day product launch, where companies often publish pricing and headline capabilities first and fill in technical documentation over the following weeks. But it is worth flagging plainly rather than filling the gaps with guesses, since a procurement team moving quickly on a voice AI contract needs to know what is actually confirmed versus what is marketing framing. ## Why Voice AI Is Suddenly the New Battleground Text-based chatbots have been the center of AI product competition for the past three years, but voice is a harder problem with a bigger commercial prize attached. Call centers, in-car assistants, telehealth triage lines, and phone-based banking all run on voice, and most of that infrastructure still relies on decades-old interactive voice response systems that frustrate callers and cost companies money in abandoned calls. A voice agent that actually sounds natural and responds inside a few hundred milliseconds is the kind of upgrade that can be sold on hard cost savings, not just novelty. That is why Mistral chose to name five or more competitors directly in its launch materials rather than speaking in generalities. The company is signaling that it wants to be evaluated as a direct alternative to OpenAI’s Realtime API, Google’s Gemini speech tools, and specialist voice vendors like ElevenLabs and Deepgram, not as a side project attached to its better-known text models. For a company whose public identity has mostly been built on large language models like Mistral Large, pivoting hard into a specialist voice product category is a notable strategic move even without a single benchmark score to back it up yet. ## Market Impact: What This Means for Enterprise Voice AI Buyers For procurement teams currently running a voice AI pilot or renewal, the Voxtral 3 pricing sheet is now a reasonable floor to negotiate against, regardless of whether a company ultimately picks Mistral. Vendors like Deepgram, ElevenLabs, and the speech arms of OpenAI and Google now have a published, named competitor undercutting them on price with specific per-minute figures attached, which tends to put downward pressure on list prices across an entire category even for buyers who never touch Mistral’s product directly. The bigger question for enterprise buyers is reliability at scale rather than headline pricing. A support line handling thousands of simultaneous calls needs consistent latency under load, graceful degradation when a connection drops, and predictable behavior across accents and background noise, none of which a launch-day pricing table can demonstrate. Expect enterprise buyers to treat Voxtral 3 as a serious evaluation candidate over the next two to three months rather than an immediate production swap, especially for anything handling regulated data like healthcare or financial calls where licensing and data residency questions need firm answers first. ## Historical Context: From Whisper to Real-Time Voice Agents Modern speech AI has moved in roughly three waves. The first wave, dominated by OpenAI’s Whisper models after their 2022 release, made reasonably accurate transcription cheap and widely accessible, but Whisper itself was never built for live, streaming conversation. The second wave added real-time capability and pushed error rates down further, with specialist vendors like Deepgram and ElevenLabs building businesses specifically around low-latency transcription and natural-sounding synthetic voices. Voxtral 3 represents a push toward a third wave: full agentic voice pipelines where transcription, reasoning, and speech generation are designed and sold together as one product rather than assembled from parts bought from three different vendors. That consolidation pattern mirrors what happened earlier in text-based AI, where standalone embedding, retrieval, and generation tools eventually got bundled into single agent frameworks. Mistral betting its voice strategy on the same bundled-pipeline approach, rather than selling a best-in-class transcription model and letting customers assemble their own stack, suggests the company has concluded that the integration itself, not any single component, is where the margin and the lock-in live. ## What the Launch Statements Say Mistral AI kept its public messaging narrow and product-focused rather than leaning on executive quotes. The company’s account on X introduced the launch plainly: “Introducing Voxtral 3: state-of-the-art Transcribe and Speech models powering Mistral’s new voice agents API,” a statement that frames the release as infrastructure for a broader voice agents product rather than a standalone model drop. A follow-up post on the same account expanded on the architecture: “Voxtral 3 also introduces a voice agent SDK, built on LiveKit. Transcribe listens, a reasoning model thinks and calls tools, and Speech replies, streamed as one pipeline for natural, intelligent conversation,” language that closely echoes the phrasing used on Mistral’s own news page. No named Mistral executive is quoted anywhere in the materials reviewed for this launch, a notable contrast with how some AI companies have used founder quotes to frame major releases. Every public statement about Voxtral 3 so far is attributed to the company account rather than an individual, which is worth keeping in mind for anyone citing this launch as reflecting a particular person’s strategic view inside Mistral. ## Five Predictions for Voice AI Through 2027 - Expect at least one of OpenAI, Google, or ElevenLabs to publish a competing price cut or a direct rebuttal benchmark within the next quarter, since Mistral named them specifically and left an opening for a response. - Expect Mistral to publish a licensing clarification and context-window specs for Voxtral 3 within weeks, following the typical pattern of same-day launches filling in documentation gaps shortly after release. - Expect call center and customer service software vendors to announce Voxtral 3 integrations faster than consumer app makers, since the per-minute pricing model maps directly onto existing call-center cost accounting. - Expect independent benchmark sites to publish third-party comparisons of Voxtral Small Transcribe 3 against Whisper large-v3 and GPT-Transcribe within one to two months, which will either validate or undercut Mistral’s launch claims. - Expect LiveKit’s role in the SDK to draw more attention from infrastructure-focused developers than the underlying models themselves, since building on shared open infrastructure rather than a proprietary transport is the more unusual strategic choice in this launch. ## The Bigger Picture for AI Product Strategy Voxtral 3 is a reminder that the AI product race is no longer only about which lab has the largest or smartest language model. Mistral is a company best known for open and efficient text models, and it chose to make its most aggressive, most specifically benchmarked competitive claims in voice, a category where it had no prior flagship product. That is a sign that speech and real-time agents have become a category worth fighting over on their own terms, separate from the general chatbot wars that dominated headlines through 2024 and 2025. For developers and buyers, the near-term task is straightforward even if the long-term competitive picture is not: test Voxtral 3 against actual call recordings or use cases, confirm licensing and data residency terms directly with Mistral before any production commitment, and watch for the inevitable round of competitive price adjustments from OpenAI, Google, ElevenLabs, and Deepgram that a launch this direct is likely to provoke. ## Frequently Asked Questions **What is Mistral Voxtral 3?** Voxtral 3 is a family of four speech AI products from Mistral AI, covering batch and streaming speech-to-text, flagship transcription with added reasoning features, and text-to-speech, plus a voice agent SDK that connects all three into one real-time pipeline. **When was Voxtral 3 released?** Mistral launched Voxtral 3 on October 20, 2026, publishing pricing, model names, and SDK details on its official news and documentation pages the same day. **How much does Voxtral 3 cost?** Pricing ranges from $0.0035 per minute for batch transcription up to $0.08 per minute for the full voice agent API with a larger reasoning model, with text-to-speech priced separately at $13 per million characters. **What is the Voxtral 3 voice agent SDK built on?** Mistral confirmed the SDK is built on LiveKit, an open real-time audio and video infrastructure layer, connecting transcription, a reasoning model, and speech generation into a single streaming pipeline. **Is Voxtral 3 open-source?** Mistral’s public governance and documentation pages do not confirm whether any Voxtral 3 model is open-weight or fully proprietary. That licensing detail remains unconfirmed as of this writing. **How does Voxtral 3 compare to OpenAI’s Realtime API and ElevenLabs?** Mistral claims Voxtral Small Transcribe 3 outperforms OpenAI’s GPT-Transcribe and Whisper large-v3, and claims Voxtral Mini TTS 3 is cheaper than ElevenLabs Flash and OpenAI’s TTS-1. No independently verified benchmark scores for these comparisons were available at launch. **What languages does Voxtral 3 support?** Mistral’s marketing materials cite more than 50 languages for real-time conversation, while separate documentation lists 35-plus languages and 50-plus dialects for expressive speech generation, alongside a narrower nine-language figure elsewhere. The exact full language list is not fully reconciled across Mistral’s own materials. **Can enterprises self-host Voxtral 3?** A third-party report describes deployment options including private cloud or on-premises infrastructure, but this claim is not corroborated by Mistral’s own official pages reviewed for this article, so self-hosting rights should be confirmed directly with Mistral before planning a deployment.