Eleven Realtime vs Voxtral 3 review: a voice agent in one box, or one you can take apart
ElevenLabs and Mistral launched voice-agent models on the same day. One gives you a whole agent behind a single setting. The other hands you the parts, with prices on each. Which one fits depends on what your phone line has to do after hello.
Verdict: pick Eleven Realtime v1 Max inside ElevenAgents if you want the shortest path to a support line that looks things up and acts, and can live with managed settings. Pick Mistral's Voxtral 3 if you need to choose the reasoning model, keep audio on your own infrastructure, or price every minute up front. Both launched October 20, 2026. Neither decides what the agent may change or when it hands the call to a person. That part is still yours.
The short verdict
On October 20, 2026, ElevenLabs and Mistral both launched new voice-agent models. ElevenLabs released Eleven Realtime, two speech-to-speech models that each act as a complete voice agent. Mistral released Voxtral 3, new speech models plus a voice-agent SDK that wires them to a reasoning model of your choice.
If you run customer support or take bookings by phone, here's our verdict. Choose Eleven Realtime v1 Max inside ElevenAgents when you want the shortest route to a phone agent that can look something up and act, and you're comfortable with ElevenLabs managing the moving parts. Choose Voxtral 3 when you need control: which model reasons, where the audio is processed, and what each minute costs before you start.
Neither product decides the two things that make or break a support line: what the agent is allowed to change, and how it hands a caller to a person. The voice is the part you'll hear. The handoff is the part your customers will remember.
Eleven Realtime: the whole agent behind one setting
ElevenLabs describes each Realtime model as a complete voice agent. Speech recognition, turn-taking, the language model and the voice are tuned together. In ElevenAgents you pick v1 Mini or v1 Max from a dropdown, and the agent keeps its prompt, voice, tools and knowledge base.
The company positions v1 Mini for natural, low-cost conversation and v1 Max for customer support, lead qualification and multi-step workflows. Its own figures put Max well ahead on tool use, completing 87.1% of tool-use tasks against 60.3% for Mini, which is why ElevenLabs recommends Max for anything that calls tools. Those are ElevenLabs' benchmarks, not an independent test. It also lists the time to first audio at 0.57 seconds for Max and 0.62 for Mini on average, again by its own measurement. Both models support more than 90 languages and the 17,000-plus voices in its library, and the API is compatible with OpenAI's Realtime API.
The catch is in the docs. In Realtime mode you can't bring a custom language model or Amazon Bedrock, workflow steps can't override the managed settings, and a Realtime agent can only transfer to or from other Realtime agents. If your HIPAA or zero-retention requirements apply, only Max is available, in the US and EU. Realtime is rolling out gradually, so it may not appear in every workspace yet.
The voice is the part you'll hear. The handoff is the part your customers will remember.
Voxtral 3: the parts, with prices on each
Mistral took the other route. Voxtral 3 is a family of speech models and an SDK, built on LiveKit, that runs listening, reasoning and speaking as one stream. A realtime transcription model hears the caller and spots when they've finished. A general model, such as Mistral Small, reasons and calls tools. Voxtral Speech 3 answers. Mistral says each turn completes in under a second, and because each stage is its own model, you can change the reasoning model, the voice or the language without rewriting the agent.
It runs on Mistral's API, with European, US and global endpoints, or on your own infrastructure if you talk to Mistral's team. Transcription covers more than 50 languages and speech more than 35. A Pro Voice option adapts a speaker's voice from about 30 minutes of audio, which matters if you want the line to sound like your business rather than a stock voice.
Mistral also published prices. The voice-agent API is listed at $0.04 a minute with Mistral Small 4 doing the reasoning and $0.08 a minute with any other reasoning model. Realtime transcription on its own is $0.008 a minute, and speech is $13 per million characters.
The catch here: phone calls over SIP are available to enterprise customers at launch, through Mistral's account team, according to its docs. Browser and app voice is open now. If your plan is "answer the main number next week," check that first.
Decision guide
What you get on launch day
| Consideration | Eleven Realtime (ElevenAgents) | Voxtral 3 (Mistral) |
|---|---|---|
| Shape | One model is the whole voice agent. | Separate listening, reasoning and speech stages. |
| Reasoning model | Managed; no custom LLM in Realtime mode. | Mistral Small or another model you choose. |
| Where it runs | ElevenLabs, with US, EU and other residency options. | Mistral API endpoints, or your own infrastructure via sales. |
| Phone lines | Phone and contact-center channels in ElevenAgents. | SIP for enterprise customers at launch. |
| Published price | Per-minute rate plus tokens; Max costs more. | $0.04 or $0.08 a minute for the agent API. |
Where both fall short
Pricing you can compare. Mistral prints per-minute rates. ElevenLabs bills the standard ElevenAgents per-minute rate plus tokens at the Realtime model's rates, with Max costing more than Mini, and the launch post doesn't give a single all-in figure. Price a real call on both before you choose.
The benchmarks are the vendors' own. Both launch posts include charts, and both are marketing. Your callers, your accents, your background noise and your tools are the only test that counts.
Neither knows your rules. A voice agent that can "check a bill" or "book an appointment" will do exactly what its tools allow. If the tool can refund, the agent can refund. We made this point in calling your AI receptionist yourself before a customer does.
The record after the call. A call ends and something has to happen: a ticket, a callback, a note in the customer's history. That's still your build. We wrote about the transcript that lost the callback for exactly this reason.
Which one fits which phone line
A small business that wants an after-hours line for bookings and simple questions, in a handful of languages, will get there fastest with ElevenAgents and Eleven Realtime. Start on Mini if the agent mostly answers questions, and move to Max once it starts calling tools.
A support team that already has a contact-center system, or a reasoning model it trusts, or a rule that audio stays in its own environment, should look hard at Voxtral 3. The separate stages are more work, and they're also what lets you swap one piece without starting over.
A regulated business, such as healthcare, should read the compliance tables before the demos. On ElevenLabs that points you to Max. On Mistral it points you to the self-deployment conversation.
How the work moves
Before the agent answers a real call
- List the top calls
The five reasons people actually phone you.
- Mark what it may change
Look up, book, cancel, refund: yes or no for each.
- Set the handoff
When it transfers, to whom, and what it says first.
- Price real calls
Each vendor's published rates at a realistic length.
- Land the record
A ticket or note someone reads the next morning.
Put this into practice before you pick
Write down the five calls your line gets most. For each, list what the agent would need to look up, what it may change, and when it must hand off to a person.
Price those five calls on both platforms, using each vendor's published rates, at a realistic length.
Decide where the call record lands when the conversation ends, and who reads it the next morning.
If you'd like a voice agent that sits on top of your booking, billing or order systems with clear limits and a proper handoff, that's AI applications and agents work we do. Talk to us.
Related work: Spin:Market.
We build alongside the people doing the work.
Bring us the problem you’re thinking about.