Key Takeaways
- Eleven Realtime is a speech-to-speech model family from ElevenLabs announced on October 20, 2026, aimed at voice agents rather than text chatbots.
- It ships in two versions: v1 Mini for speed and v1 Max for higher-precision spoken reasoning, both accessed through ElevenAgents or the ElevenAPI.
- ElevenLabs says the model starts talking in about half a second and can answer spoken reasoning questions correctly, combining low latency with accuracy.
- There is no dedicated product page yet, so the official site, elevenlabs.io, is the place to check for access details, limits and pricing.
Quick Facts
| Type | Speech-to-speech AI model family for voice agents |
|---|---|
| Devices | Cloud service used through a browser or API; microphone needed |
| Price | Not confirmed; check elevenlabs.io for current plans |
| Languages | Reported support for over 90 languages |
| Highlight | Replies in under half a second, with spoken reasoning |
| Learning curve | Low for chatting, moderate for API integration |
| Category | AI & Models |
| Creator | ElevenLabs |
| Official site | elevenlabs.io ↗ |
| Heat | 🔥 222.3K Heat |
| Last updated |
What ElevenLabs Built With Eleven Realtime
Eleven Realtime is a speech-to-speech model family from ElevenLabs, announced on October 20, 2026. Instead of converting your voice to text, generating a text answer and then converting that back to audio, it works directly from speech to speech. That design is what makes the low latency possible, and it is why ElevenLabs describes it as a new class of models rather than another text model with a voice bolted on.
The family currently has two members: v1 Mini and v1 Max. ElevenLabs positions v1 Mini as the fast option for everyday conversation, while v1 Max targets higher-precision reasoning when the spoken question is harder. Both are meant for voice agents, the kind of assistant you call or talk to and expect to answer naturally, not a text chatbot that happens to read its replies aloud.
The company's announcement frames the core problem plainly: voice agents usually force you to choose between quick replies and precise answers. Eleven Realtime is the attempt to remove that trade-off. ElevenLabs says it starts talking in about half a second and still gets spoken reasoning questions right, which is the combination most voice products struggle to deliver.
- Speech in, speech out
The model consumes audio and produces audio directly, which cuts the delay that comes from transcribing and re-synthesizing every turn.
- Two versions, two priorities
v1 Mini leans toward speed for casual back-and-forth, while v1 Max is aimed at questions that need careful reasoning.
- Built for agents
Eleven Realtime is designed for conversational voice agents, not for one-off text generation or batch transcription jobs.
- Broad language and voice coverage
Reports point to support for over 90 languages and a catalog of roughly 17,500 voices, useful if you serve an international audience.
How the Half-Second Response Actually Works
The headline claim is latency: ElevenLabs says the listen, think and answer cycle finishes in under half a second, and the CEO has described the whole process as taking less than half a second. In a live conversation, that gap is the difference between an agent that feels present and one that feels like it is waiting for a turn to end. Short pauses are what make spoken dialogue feel awkward, so this number matters more than raw model size.
The second half of the promise is accuracy on spoken reasoning. Fast models often answer quickly but shallowly, guessing at the gist of a question. Eleven Realtime is presented as handling reasoning questions correctly while keeping the fast start, which is why the two-version split exists: you pick the trade-off you want instead of accepting one fixed point.
Access runs through ElevenLabs' existing surfaces. You reach the model through ElevenAgents or the ElevenAPI, so the same account and tooling you may already use for other ElevenLabs products apply here. That also means the practical limits, concurrency rules and pricing are tied to your ElevenLabs plan rather than to a separate standalone app.
- Latency is the product
Under half a second to first speech is the feature being sold; everything else is judged against whether that speed survives real network conditions.
- Reasoning without the wait
The claim is not just fast talk but correct answers to spoken reasoning prompts, the harder half of the problem.
- One account, two entry points
ElevenAgents suits hands-on testing, while the ElevenAPI is the path for putting the model inside your own product.
- Unconfirmed details
Exact latency under load, rate limits and pricing have not been published in detail, so verify them on the official site before you commit.
Setting Up a Voice Agent With Eleven Realtime
Start with an account on the official site, elevenlabs.io, then open ElevenAgents if you want to try the model without writing code, or the ElevenAPI if you are building an integration. From there you choose a version: v1 Mini when you want snappy conversational replies, v1 Max when your use case involves questions that need real reasoning before the answer.
Test in a quiet room with microphone access enabled in your browser. Background noise, echo and a poor headset will distort your impression of both latency and accuracy, and those are exactly the two things you are evaluating. Run the same set of questions against both versions so the comparison is fair, and note where v1 Mini starts to stumble and whether v1 Max fixes it.
If you are integrating through the API, treat the model as a streaming audio component rather than a request-response endpoint. Plan for handling partial audio, interruptions and turn-taking in your client, and decide up front how your agent should behave when the user talks over it. These interaction details usually matter more to perceived quality than any single benchmark number.
- Pick the version deliberately
Use v1 Mini for chat-style agents and v1 Max for support, tutoring or anything where a wrong quick answer costs more than a short pause.
- Control your test environment
A quiet room, a decent microphone and a stable connection give you a realistic read on the half-second claim.
- Compare on the same script
Ask both versions identical questions and score latency and correctness separately, so you can see which one fails and why.
- Design for interruption
Real callers cut in; decide how your agent stops, listens and resumes before you ship it to users.
Where Eleven Realtime Fits and Where It Does Not
The obvious fit is anything that behaves like a phone call: customer support lines, booking assistants, language practice partners, in-app helpers and interactive characters. In those settings, a sub-second reply is not a luxury, it is the difference between a conversation and a walkie-talkie exchange. The reported support for over 90 languages and a large voice catalog also makes it practical for products that serve many regions at once.
Compared with the common alternative of chaining a speech recognizer, a text model and a text-to-speech engine, Eleven Realtime removes two handoffs and the latency they add. The trade-off is control: a pipeline lets you inspect and edit the transcript at each stage, while a direct speech-to-speech model gives you less visibility into what happened between hearing and answering. If auditability or transcript editing is a hard requirement, that difference matters.
It is also not a general-purpose text model, a music generator or a transcription service. And because there is no dedicated product page yet, several practical questions remain open: exact pricing, rate limits, data handling and whether the model will be offered outside the ElevenLabs platform. Anyone planning a production rollout should confirm those points on the official site rather than assume them.
- Strong fit
Live support, scheduling, tutoring and companion agents where natural turn-taking and broad language coverage matter most.
- Weaker fit
Workflows that need an editable transcript or a full audit trail at every step are better served by a staged pipeline.
- Not the same tool
Eleven Realtime is for real-time spoken dialogue, not text generation, transcription or voice cloning as a standalone task.
- Open questions
Pricing, limits and platform availability are unconfirmed; check elevenlabs.io for the current state before building around it.
Step-by-Step Guide
Create your ElevenLabs account
Go to the official site, elevenlabs.io, and sign up or log in. This single account gives you access to both ElevenAgents and the ElevenAPI, and it is where any plan limits for Eleven Realtime will appear.
Open ElevenAgents or the ElevenAPI
Use ElevenAgents for a no-code trial of the voice agent experience, or the ElevenAPI if you are wiring the model into your own application. Both are the supported routes into Eleven Realtime.
Choose v1 Mini or v1 Max
Select v1 Mini when you want the fastest conversational replies, and v1 Max when your questions require careful reasoning. You can switch between them to compare on the same prompts.
Grant microphone access and test
Allow microphone permissions in your browser and run your session in a quiet room. Speak naturally, include one or two reasoning questions, and judge both how fast the reply starts and whether the answer is right.
Compare results and plan integration
Score latency and accuracy for each version, then decide which one fits your use case. If you are going to production, confirm pricing, limits and data handling on the official site before you build.
Frequently Asked Questions
What exactly is Eleven Realtime?
It is a speech-to-speech model family from ElevenLabs, announced on October 20, 2026, and designed for voice agents. Rather than transcribing speech to text and back again, it works directly from audio input to audio output. It comes in two versions, v1 Mini and v1 Max, and is accessed through ElevenAgents or the ElevenAPI.
How fast is it, and is the half-second claim reliable?
ElevenLabs says the model starts talking in about half a second, and the CEO has described the full listen, think, answer cycle as taking under half a second. Real-world latency will depend on your connection, device and how the service is loaded at the time. Test it yourself in your own environment before relying on that figure.
What is the difference between v1 Mini and v1 Max?
v1 Mini prioritizes speed, making it the better choice for casual chat and simple requests where an instant reply matters most. v1 Max targets higher-precision reasoning, so it suits questions where a fast but wrong answer is worse than a slightly slower correct one. Running the same prompts through both is the quickest way to see the difference.
How do I access Eleven Realtime, and what does it cost?
Access is through ElevenLabs: sign up on the official site, elevenlabs.io, then use ElevenAgents or the ElevenAPI. Pricing and usage limits for Eleven Realtime have not been published in detail, and there is no dedicated product page yet, so check the official site for current plans and restrictions before you commit.
Which languages and voices does it support?
Reports indicate support for over 90 languages and a catalog of roughly 17,500 voices, which would make it unusually broad for a real-time speech model. Those figures come from early coverage rather than a published specification sheet, so confirm the exact language and voice list on the official site if your product depends on a specific one.
