Realtime voice agent models have just had their most important week yet. ElevenLabs Realtime, launched on 20 October 2026, is a family of speech-to-speech models that starts speaking 570 milliseconds after you finish talking — fast enough that a phone call with an AI agent finally feels like a conversation instead of a transaction. Our verdict up front: if you are building a voice agent for support, sales, or scheduling, ElevenLabs Realtime is now the model to beat on speed and spoken reasoning, but it is not the right pick for every team. The cheaper Mini model is too weak at tool use for serious workflows, and the all-in-one design trades away the control that a hand-built pipeline gives you.
What ElevenLabs Realtime actually is
Until now, building a voice agent meant wiring together at least four separate pieces: a speech recogniser to transcribe what the caller said, a language model to decide what to say back, turn-taking logic to judge when to speak and when to wait, and a voice synthesiser to speak the answer. Every hand-off between those pieces added delay, and every hand-off was a place where the conversation could go wrong — the agent talking over the caller, or freezing for a second while it thought.
ElevenLabs Realtime collapses all of that into a single model. Audio goes in and audio comes out, with recognition, reasoning, turn-taking, and voice generation tuned together rather than chained. That is the whole pitch, and it is a real architectural change rather than a marketing rebrand: when one model owns the entire turn, it can decide to interrupt, pause, or murmur a backchannel like “mm-hmm” the way a human would, because it is not waiting for a transcript to travel between systems.
There are two models in the family. Eleven Realtime v1 Mini is built for natural conversation — assistants, tutors, and interactive characters where the flow of talk matters most. Eleven Realtime v1 Max is built for complex work — customer support, lead qualification, and multi-step tasks that require calling tools. Both support more than 90 languages and can draw on ElevenLabs library of over 17,000 voices, so a business serving customers in several languages does not need a different voice stack per market.
Mini vs Max: the numbers that matter
ElevenLabs reports its benchmark figures from Artificial Analysis, an independent evaluation firm, so these numbers are vendor-reported but independently run. Read them as a shortlist filter, not a guarantee — run your own test calls before committing to any voice model.
- Time to first audio: 570 milliseconds for Max, 620 milliseconds for Mini. Both sit near the 600-millisecond threshold researchers use for when an interaction starts to feel conversational rather than transactional.
- Spoken reasoning (Big Bench Audio, 1,000 spoken questions): Max answers 98.8% correctly, Mini 75.9%. That gap is the single biggest reason to pick Max for anything beyond chat.
- Task success (tool use): Max completes 87.1% of tasks, Mini only 60.3%. A Mini agent will sound lovely and then fail to actually book the appointment or check the order.
- Turn-taking (Full-Duplex-Bench): Mini actually wins here, 95.8% against 94.1% for Max, because the smaller model is tuned for the rhythm of conversation — pauses, interruptions, and backchannels.
- Blind preference (Speech Agent Arena): Max scores 998 and Mini 988 in live comparisons where listeners do not know which model they heard.
The practical reading is simple. Mini is the conversation model and Max is the work model. If your agent mostly talks — a tutor, a character, a friendly front door — Mini gives you the better rhythm at a lower cost. If your agent has to do things — look up an order, qualify a lead, book a slot — Max is the only one of the two that clears the bar, and the price difference is worth paying.
Realtime voice agent models and the latency race
ElevenLabs did not launch into an empty market. The same week, Mistral launched Voxtral 3, a self-hostable voice stack aimed at regulated industries that cannot send call audio to US clouds. Grok Voice 3.0 answers in 810 milliseconds with a higher task-success rate (94.8%) than Max, and Gemini Live and OpenAI GPT-Live trail on speed at over a second to first audio. ElevenLabs claim — that no model on the Artificial Analysis speech-to-speech leaderboard is both faster and more accurate than Max on spoken reasoning — held at launch, but this leaderboard moves weekly. Speed is the battleground right now because every vendor has realised the same thing: a voice agent that pauses for a second feels broken, however smart its answers are.
What ElevenLabs is really selling is the removal of integration work. The company line is that you should not have to become a configuration expert to build a voice agent that sounds natural, and for most businesses that is true. A small team that would have spent weeks tuning a pipeline can now pick a model from a dropdown in ElevenAgents, the hosted platform, and get phone, web widget, WhatsApp, and contact-centre channels, plus a knowledge base, guardrails, scripted scenario testing, and production monitoring, in one place.
Pricing: what it actually costs to run
Here is the honest gap in the launch: ElevenLabs published no separate Realtime price list. Reporting puts the ElevenAgents rate at about $0.08 per minute plus language-model token charges, with Max costing more than Mini. That per-minute figure is only part of the bill. Token charges scale with how much your agent talks and how many tools it calls, so a support agent that chats for six minutes and checks three systems costs materially more than a two-minute booking agent. Before you budget, run a week of real calls through the platform and price the invoice, not the headline rate. For a business handling a few hundred calls a month, the total is typically a rounding error next to one support hire. For a high-volume contact centre, per-minute pricing deserves a spreadsheet before it deserves a signature.
Who it is for — and who should wait
ElevenLabs Realtime is a strong fit for support and sales teams that lose leads to missed calls, for appointment-based businesses (clinics, salons, trades, and professional services) where booking is the whole job, and for education and coaching products where a tutor that listens and responds naturally is the product. It is also a good fit for teams already built on OpenAI Realtime API: the ElevenLabs API is compatible with it, so switching means changing the endpoint and model name rather than rebuilding the integration — though note the API runs over WebSocket only, so a WebRTC integration has to move transport first.
Wait if you need full control over individual components — for example, a custom recogniser for a niche accent, or a specific language model for compliance reasons. Wait if you need on-premise or fully self-hosted voice processing; this is a hosted model family, and Mistral Voxtral 3 is the more interesting option there. And wait if your use case is text-first with voice as an add-on: a cheaper cascade pipeline will do that job without paying per-minute voice-model rates.
Watch-outs and honest limitations
- You lose component control. One model means you cannot swap in a better transcriber or a different voice engine. If ElevenLabs changes the model, your agent changes with it.
- No transcript layer by default. The single-model design removes the text transcript that most enterprise safety and compliance tooling is built around. Check how conversation logs, which are searchable and auditable in ElevenAgents, map to your compliance needs before you deploy.
- Mini has compliance gaps. Mini does not support Zero Retention or HIPAA modes; Max supports those in the US and EU. Max is also not available in the India or Singapore data-residency environments, which matters if your data-residency policy names those regions.
- Rollout is gradual. Realtime is rolling out workspace by workspace, so you may not get access on day one, and the separate API does not include ElevenAgents channels or knowledge-base features.
- Vendor benchmarks are a starting point. All figures above are ElevenLabs-reported results from Artificial Analysis tests. Independent replication had not landed at launch. Trust, but verify with your own calls.
How it compares to the alternatives
Against OpenAI Realtime API, ElevenLabs wins on speed and voice quality and loses on ecosystem: OpenAI is already inside many teams stacks. Against Vapi and similar turnkey telephony platforms, ElevenLabs wins on voice naturalness and loses on telephony depth — Vapi is built around call routing first. Against Deepgram and AssemblyAI, the comparison is not direct: those are best-in-class components for teams that want to build their own pipeline, while ElevenLabs is the argument for not building one. And against Mistral Voxtral 3, the choice is philosophical: ElevenLabs is polished and hosted, Voxtral is open and self-hostable. Pick the philosophy that matches your compliance posture, then pick the model.
From our experience
MTD Technologies has spent 9+ years and 300+ projects building the kind of systems voice agents have to talk to — CRMs, booking tools, support desks, and order systems. The pattern we see again and again is that the voice model is never the hard part. The hard part is what happens after the caller says something: the agent has to check a real calendar, a real order status, or a real CRM record, and it has to do it without inventing an answer when the system is slow or the record is missing. That is where most voice-agent pilots quietly fail — not on latency, but on the last mile of integration and on guardrails that stop the agent from confirming things it has not verified.
Our advice to clients is always the same: pilot with one narrow job (missed-call capture, or appointment booking, or order-status checks), measure the completion rate and the cost per completed call for two weeks, and only then widen the scope. ElevenLabs Realtime makes the pilot cheaper to start, because you skip the pipeline build. It does not remove the integration work, the testing, or the human handoff design — and any vendor who tells you otherwise is selling you a demo, not a deployment. If you want help scoping a voice-agent pilot properly, our AI integration and automation team does exactly this work, and you can browse more tools like this one in our AI Tools Directory.
Frequently asked questions
What is ElevenLabs Realtime?
ElevenLabs Realtime is a family of speech-to-speech models for building voice agents. Instead of chaining a transcriber, a language model, and a voice synthesiser, each model handles the whole conversation — audio in, spoken answer out — in a single system.
How fast is ElevenLabs Realtime?
Eleven Realtime v1 Max starts speaking in 570 milliseconds on average, and v1 Mini in 620 milliseconds, according to Artificial Analysis evaluations reported by ElevenLabs. Both are near the threshold where a voice interaction starts to feel conversational.
Should I choose Mini or Max?
Choose Mini for conversation-first uses like assistants, tutors, and characters, where turn-taking matters most. Choose Max for support, lead qualification, and multi-step work, because Mini only completes 60.3% of tool-use tasks against 87.1% for Max.
How much does ElevenLabs Realtime cost?
ElevenLabs has not published a separate Realtime price list. Reporting puts the ElevenAgents rate at about $0.08 per minute plus language-model token charges, with Max costing more than Mini. Run a week of real calls to price your actual invoice before budgeting at scale.
Can I switch from OpenAI Realtime API?
Yes, in most cases. The ElevenLabs Realtime API is compatible with OpenAI Realtime API, so switching means changing the endpoint and model name. One catch: the API runs over WebSocket only, so a WebRTC integration has to move transport first.
The bottom line on realtime voice agent models is that latency was the last excuse for a voice agent that feels robotic, and ElevenLabs Realtime just removed it. What remains is the unglamorous work that decides whether a voice agent makes money: real integrations, real guardrails, and a human handoff that works. Get those right and 570 milliseconds is a genuine competitive edge. If you would like help scoping that pilot, talk to MTD Technologies — we build the systems behind the voice, not just the voice itself.