On October 20, 2026, the AI voice race changed in a single day. Mistral Voxtral 3 arrived as a full voice-agent stack that listens, reasons, calls tools, and speaks in one streaming pipeline — and within hours, ElevenLabs fired back with Eleven Realtime v1, a pair of speech-to-speech models built for live conversation. Two completely different philosophies, launched on the same day, both claiming to power the voice agents that will answer your next support call.
If you build products, run a business, or just want to know which voice AI to bet on, this comparison matters. I have pulled the launch details, the benchmark claims, and the pricing for both platforms so you can see exactly where Voxtral 3 and Eleven Realtime v1 win — and where their marketing needs a reality check.
Table of Contents
- Why October 20, 2026 changed the AI voice agent race
- What is Mistral Voxtral 3?
- What is ElevenLabs Eleven Realtime v1?
- Voxtral 3 vs Eleven Realtime: head-to-head comparison
- Pricing breakdown: what an AI voice agent actually costs
- How a modern AI voice agent actually works
- Which should you choose?
- The honest caveats (read before you buy)
- How to get started today
- Frequently asked questions
- The bottom line
Why October 20, 2026 changed the AI voice agent race
Voice AI had been moving fast all year, but October 20, 2026 was the day the market split into two camps. On one side, Mistral AI — the Paris lab known for open-weight models and its Mistral Large 4 release — launched Voxtral 3, a full-stack voice agent system with separate, swappable parts: transcription models, speech models, and an SDK that ties them to a reasoning model in a single streaming pipeline. On the other side, ElevenLabs — already a giant in synthetic voices through its ElevenLabs v4 line — launched Eleven Realtime v1, two end-to-end speech-to-speech models that absorb the whole pipeline into one managed service.
The timing is not a coincidence. Voice agents are moving from demos to deployment: contact centers, appointment booking, lead qualification, and customer support are all being rebuilt around machines that can hold a phone conversation. Artificial Analysis now runs dedicated speech-to-speech and speech-to-text leaderboards, and Voxtral 3 entered the STT rankings at number three on day one, while Eleven Realtime v1 Max landed near the top-left corner of the speech-to-speech board — the spot that means both fast and accurate.
For US businesses, the stakes are practical. A voice agent that can handle a support call for a few cents changes the math on staffing, after-hours coverage, and missed calls. But the two stacks make different promises: Mistral sells control, data sovereignty, and per-component pricing, while ElevenLabs sells simplicity — one model, one API, no assembly required. Let us look at each one properly.

What is Mistral Voxtral 3?
Mistral Voxtral 3 is Mistral AI’s third-generation voice family, and it is a deliberate break from the “one model does everything” approach. Instead of a single black box, Mistral ships three pieces you can combine: transcription models that turn speech into text, speech models that turn text into speech, and a voice-agent SDK — built on the open-source LiveKit framework — that streams listening, reasoning, tool calls, and speaking as one real-time pipeline.
The pitch is aimed squarely at businesses that need voice agents to do work during a call: check an invoice, book an appointment, pull up a record, and confirm the result before the caller hangs up. Everything is available now through Mistral’s API (EU, US, and global endpoints), and — crucially — companies can also deploy the whole stack on their own infrastructure or in a private cloud. That self-hosting option is Mistral’s sharpest differentiator against US-hosted competitors.
Voxtral 3 transcription models: 50+ languages
The listening side of Voxtral 3 comes in three flavors:
- Voxtral Small Transcribe 3 — the flagship transcription model, covering more than 50 languages including all 24 official EU languages. It includes speaker diarization (knowing who said what), context biasing for domain-specific vocabulary, and end-of-turn detection so the agent knows when you have finished talking. Pricing is $0.0040 per minute.
- Voxtral Mini Transcribe 3 — a smaller 4-billion-parameter model for batch transcription at $0.0035 per minute.
- Voxtral Mini Transcribe Realtime 3 — built specifically for live agents at $0.008 per minute, with what Mistral says is sub-500-millisecond latency.
On accuracy, Mistral published a 45-language FLEURS chart showing error rates of 4.1% (language specified) and 4.0% (language auto-detected) — matching Microsoft AI Transcribe 2 and, per Mistral’s chart, ahead of Gemini 3.5 Transcribe, Scribe v2, GPT Transcribe, and Whisper large-v3. Artificial Analysis placed Voxtral Small Transcribe 3 at number three on its speech-to-text leaderboard at launch. These are company-reported numbers, so treat them as directional rather than gospel, but the third-party leaderboard placement adds real weight.

Voxtral 3 speech models and Pro Voice
The speaking side is Voxtral Speech 3 (also called Voxtral Mini TTS 3), which supports 35+ languages and 50+ dialects. Rather than reading text flat, it adjusts emotion, pacing, and emphasis to fit the conversation — the difference between a robot reading a script and a voice that sounds like it is actually listening. Pricing is $13 per 1 million characters, which Mistral says undercuts comparable models from OpenAI, Deepgram, ElevenLabs Flash, and Gemini Flash.
The eye-catching addition is Pro Voice: feed the system up to 30 minutes of reference audio and it adapts a speaker’s vocal character, emotional expression, and style across languages. API access for this is expected to cost roughly $0.001 per minute — about six cents for a full hour of processing. That is a remarkable price for cross-language voice cloning, and it is clearly aimed at media, localization, and content teams as much as at call centers.
The LiveKit-based voice agent SDK
The piece that turns models into agents is the Voxtral 3 voice agent SDK, built on LiveKit. It combines real-time transcription, a reasoning model, tool calling, and streamed speech output into one pipeline. In practice, that means an agent can hear you, think, hit an external system (a booking API, a CRM, a billing database), and keep talking — all inside the same conversation, without the awkward pauses that plague stitched-together voice bots.
Mistral is also framing the toolkit around workflow automation rather than just conversation: voice agents that kick off tasks, not only answer questions. And the deployment story stays consistent — run it through Mistral’s hosted APIs or keep everything on your own servers, which matters for regulated industries and for European companies navigating data-sovereignty rules.

What is ElevenLabs Eleven Realtime v1?
ElevenLabs Eleven Realtime v1 takes the opposite architectural bet. Where Voxtral 3 gives you separate components to assemble, ElevenLabs absorbs the entire voice-agent stack — speech recognition, turn-taking, voice activity detection, reasoning, and speech generation — into two end-to-end speech-to-speech models. You do not pick a transcriber and a TTS engine; you pick a model and talk to it.
The launch includes two variants for different workloads:
Eleven Realtime v1 Max: accuracy first
Eleven Realtime v1 Max is the heavyweight. ElevenLabs reports 98.8% on the Big Bench Audio benchmark — which tests reasoning through spoken questions and answers — with an average time to first audio of 0.57 seconds. The company says no model on the Artificial Analysis leaderboard is currently both faster and more accurate. Max is aimed at demanding tasks: customer-support bots, lead qualification, and multi-step workflows where a wrong answer costs real money.
Eleven Realtime v1 Mini: conversation first
Eleven Realtime v1 Mini is the cheaper, chattier sibling, built for natural everyday conversation — AI assistants, teachers, interactive characters — where cost matters more than deep reasoning. On Big Bench Audio it scored 75.9% with 0.62 seconds to first audio, but here is the interesting part: on Full-Duplex-Bench, which measures conversational dynamics like pauses, interruptions, turn-taking, and backchannels, Mini actually beat Max, 95.8% to 94.1%. Mini is also the one that can tell when a speaker has finished versus when they are just saying “mm-hmm” — the small human details that make a voice agent feel natural instead of mechanical. These comparative figures come from Artificial Analysis evaluations across three public benchmarks, which is the closest thing this launch has to independent verification.
ElevenAgents and the OpenAI-compatible API
Both models are available through the ElevenAgents platform and an OpenAI-compatible Realtime API — a deliberate choice that makes it cheap for developers already on OpenAI’s voice stack to test ElevenLabs without rewriting their integration. They support 90+ languages and can draw on ElevenLabs’ library of more than 17,000 voices. Pricing is per minute of audio plus token usage, with Max priced higher than Mini; ElevenLabs points developers to its pricing page for the exact numbers, which were not included in the launch materials.
One practical note from the launch coverage: developers connecting through WebRTC will need to move to WebSockets — worth confirming with your developer before planning a migration. And as Tech Times pointed out, the managed-service design means you get simplicity at a cost: no text transcript of the conversation and no ability to swap individual components, which matters for enterprises whose safety and compliance tooling is built around transcripts.
Voxtral 3 vs Eleven Realtime: head-to-head comparison
Here is how the two stacks compare on the dimensions that actually decide a purchase:
| Dimension | Mistral Voxtral 3 | ElevenLabs Eleven Realtime v1 |
|---|---|---|
| Architecture | Modular: separate transcribe, speech, and agent-SDK components you assemble | End-to-end: two speech-to-speech models absorb the whole pipeline |
| Languages | 50+ (transcribe) / 35+ languages and 50+ dialects (speech) | 90+ languages, 17,000+ voices |
| Latency | Sub-500ms for the realtime transcription model | 0.57s (Max) / 0.62s (Mini) time to first audio |
| Headline benchmark | #3 on Artificial Analysis STT board; 4.0–4.1% FLEURS error rate (company-reported) | 98.8% Big Bench Audio (Max, company-reported); Mini wins Full-Duplex-Bench 95.8% |
| Starting price | $0.0035/min transcription; $13 per 1M chars speech; ~$0.001/min Pro Voice | Per-minute + token pricing (Max > Mini); check pricing page |
| Deployment | EU/US/global APIs or self-hosted on your own infrastructure | Hosted via ElevenAgents / OpenAI-compatible Realtime API |
| Voice cloning | Pro Voice: 30-min reference, cross-language adaptation | Full ElevenLabs voice library; 17,000+ existing voices |
| Transcript access | Yes — you control each component, so transcripts are available | No — managed service, no transcript exposed |
| Best for | Regulated industries, EU data sovereignty, custom pipelines | Fastest path to a natural-sounding agent; OpenAI-stack teams |
The core trade-off is control versus convenience. Voxtral 3 is a toolkit: you can swap the reasoning model, keep data on your own servers, and see every intermediate step. Eleven Realtime v1 is an appliance: simpler to ship, harder to inspect. Neither is wrong — they are built for different buyers.

Pricing breakdown: what an AI voice agent actually costs
Voice-agent pricing is a stack, not a single number: you pay for listening (transcription), thinking (the reasoning model), and speaking (TTS or speech-to-speech). Here is what the two launches actually published in USD:
- Voxtral 3 transcription: $0.0035/min (Mini, batch), $0.0040/min (Small), $0.008/min (Mini Realtime 3).
- Voxtral 3 speech: $13 per 1 million characters — roughly a few dollars per hour of spoken output depending on speaking rate.
- Voxtral 3 Pro Voice: ~$0.001/min of API processing — about $0.06 per hour.
- Eleven Realtime v1: per-minute-of-audio plus token pricing, Max tier above Mini. Exact rates live on ElevenLabs’ pricing page.
To put that in business terms: an hour of AI-handled phone time on the Voxtral 3 stack can cost well under a dollar in model fees for the speech components — the transcription side of an hour-long call runs about $0.21–$0.48 depending on the model, and speech output adds a few dollars more. Compare that with a human support agent’s hourly cost and the economics of after-hours coverage start to look obvious. The real bill, of course, includes the reasoning model in the middle and whatever your telephony provider charges — but the voice layer itself has never been cheaper.
How a modern AI voice agent actually works
Whether you pick Voxtral 3‘s modular stack or Eleven Realtime’s single model, the conversation follows the same loop:
- Listen: your voice is captured and converted to text (transcription), with the system detecting when you stop speaking (end-of-turn detection) versus when you are just pausing.
- Think: a language model reads the transcript plus conversation history and decides what to do — answer directly or call a tool.
- Act: the agent can hit external systems mid-call: check a booking, look up an invoice, update a CRM record.
- Speak: the response is converted back to speech and streamed to you, with natural pacing, pauses, and interruptions handled in real time.
The hard part was never any single step — it was doing all four fast enough that the caller never notices. Sub-second latency (under 500ms for Voxtral 3‘s realtime transcription, 0.57 seconds for Eleven Realtime Max’s first audio) is the threshold where a conversation stops feeling like talking to a machine and starts feeling like talking to a person. Both launches clear that bar, which is why this particular day mattered: the technology crossed from “impressive demo” to “deployable product” for both vendors at once.

Which should you choose?
The honest answer depends on what you value most:
Choose Mistral Voxtral 3 if you need data to stay in specific jurisdictions (the EU-hosting and self-deploy options are unmatched at this price), you want to swap components or inspect transcripts for compliance, you are already in Mistral’s ecosystem, or you want the cheapest published per-unit pricing for transcription and speech. The modular design also ages better: when a better reasoning model appears, you swap one piece instead of rebuilding.
Choose ElevenLabs Eleven Realtime v1 if you want the shortest path from idea to a natural-sounding agent, you are already building on OpenAI’s Realtime API (the compatibility makes testing nearly free), you need the 17,000-voice library or 90+ language coverage, or your priority is conversational feel — Mini’s Full-Duplex-Bench win suggests it handles the messy human parts of conversation (interruptions, backchannels) better than anything else measured so far.
For most small US businesses evaluating their first voice agent, the practical advice from the launch coverage holds: the technical switching cost is low, so run a real test. Point both at a handful of actual customer calls, measure resolution rate and caller satisfaction, and let the numbers decide. Neither vendor’s benchmark sheet will tell you how your customers react to a machine answering their call.
The honest caveats (read before you buy)
Both launches deserve scrutiny alongside the applause:
- Benchmarks are vendor-reported. Mistral’s FLEURS chart and ElevenLabs’ 98.8% Big Bench Audio figure come from the companies’ own launch materials. The exceptions worth noting: Artificial Analysis independently placed Voxtral Small Transcribe 3 at #3 on its STT leaderboard, and its three-benchmark evaluation of Max vs Mini is the closest to independent verification either launch has. Independent, head-to-head testing of Voxtral 3 against Eleven Realtime v1 does not exist yet.
- ElevenLabs trades away transparency. The managed speech-to-speech design means no conversation transcript and no component swaps. If your compliance, QA, or safety tooling depends on reading what was said, that is a real architectural limitation, not a missing feature.
- Migration details matter. ElevenLabs developers on WebRTC need to move to WebSockets. Mistral’s self-hosting story is powerful but means you operate the infrastructure. Budget integration time, not just model fees.
- Disclosure is a legal question, not a technical one. Voice agents handle customers’ voices and personal data. Tell callers they are speaking with an AI, and review what gets recorded and where it is stored — especially under US state privacy laws and, for EU callers, GDPR.
- Language counts are not quality guarantees. “50+ languages” or “90+ languages” describes coverage, not equal quality in each. If your customers speak a less common language, test in that language specifically.
How to get started today
With Voxtral 3: sign up for Mistral’s API, where the transcription and speech models are available now across EU, US, and global endpoints. The voice agent SDK documentation walks through wiring transcription, a reasoning model, tool calls, and speech into the LiveKit streaming pipeline. If you need self-hosting, talk to Mistral about private deployment — that path exists but is not self-serve in the same way.
With Eleven Realtime v1: the fastest route is the ElevenAgents platform, which handles the telephony and orchestration for you. Developers can also hit the OpenAI-compatible Realtime API directly — if you already have code against OpenAI’s voice stack, the endpoint and model name are nearly the only things that change. Start with Mini for conversational use cases and graduate to Max where accuracy and tool use justify the higher per-minute cost.
With either: before committing, define what a successful call looks like for your business (booking completed? issue resolved? lead qualified?), run a small pilot with real traffic, and measure against your current baseline. The vendors will happily sell you accuracy percentages; your call logs will tell you the truth.
Frequently asked questions
What is Mistral Voxtral 3?
Voxtral 3 is Mistral AI’s voice-agent stack, launched October 20, 2026. It combines transcription models (50+ languages), speech models (35+ languages, 50+ dialects), and a LiveKit-based voice agent SDK into one real-time streaming pipeline, with EU/US/global API hosting or self-deployment on your own infrastructure.
What is ElevenLabs Eleven Realtime v1?
Eleven Realtime v1 is ElevenLabs’ pair of end-to-end speech-to-speech models for voice agents, also launched October 20, 2026. v1 Max targets accuracy (98.8% on Big Bench Audio, 0.57s response), while v1 Mini targets natural, low-cost conversation (95.8% on Full-Duplex-Bench). Both support 90+ languages and work through ElevenAgents or an OpenAI-compatible Realtime API.
Voxtral 3 vs Eleven Realtime: which is better?
They optimize for different things. Voxtral 3 wins on control, data sovereignty, component transparency, and published per-unit pricing. Eleven Realtime v1 wins on simplicity, conversational naturalness, and the 17,000-voice library. For regulated or EU-focused deployments, Mistral’s self-hosting is decisive; for the fastest path to a natural-sounding agent, ElevenLabs has the edge.
How much does an AI voice agent cost in 2026?
Model fees for the voice layer have fallen below a dollar per hour of call time on the Voxtral 3 stack ($0.0035–$0.008/min transcription, $13 per 1M characters of speech). ElevenLabs prices Eleven Realtime v1 per minute of audio plus tokens, with Max above Mini. Add the reasoning model and telephony costs for the full bill — but the voice components are now the cheapest part of the stack.
How many languages do Voxtral 3 and Eleven Realtime support?
Voxtral 3 transcription covers 50+ languages (including all 24 official EU languages) and speech covers 35+ languages with 50+ dialects. Eleven Realtime v1 supports 90+ languages. Test quality in your specific languages before deploying — coverage counts do not guarantee equal quality everywhere.
The bottom line
October 20, 2026 will be remembered as the day voice AI grew up. Mistral Voxtral 3 proved that an open, modular, self-hostable voice stack can match the best closed systems on accuracy while undercutting them on price — and its EU-hosting and private-deployment options give it a genuine moat with regulated and European buyers. ElevenLabs Eleven Realtime v1 proved the opposite thesis just as convincingly: absorb the whole pipeline into one model, make it sound human down to the “mm-hmm”s, and developers will come.
The winner is not decided by a benchmark table. It will be decided by whose agents actually resolve calls, book appointments, and keep customers happy at a price businesses can afford — measured in real deployments over the coming months. What is certain is that the era of the awkward phone bot is over. The next voice that answers your call may well be running on Voxtral 3 or Eleven Realtime v1, and after October 20, you probably will not be able to tell.
Sources: The Decoder’s Voxtral 3 launch report, The Decoder’s Eleven Realtime report, TestingCatalog, Analytics India Magazine, Tech Times, Mistral’s Voxtral 3 announcement.
