Introduction

ElevenLabs has a new kind of voice model. Eleven Realtime, announced on October 20 in the ElevenLabs launch post, is a family of speech-to-speech models where each one is a complete voice agent: it listens, decides when you're done talking, thinks, and answers in an ElevenLabs voice, all inside one system. There are two models, v1 Mini and v1 Max, and both are live today in ElevenAgents and through a new Realtime API that follows OpenAI's Realtime spec. This is a closed, hosted release with no open weights, priced at $0.08 a minute plus LLM usage on the API pricing page.

What Shipped

Two models, two jobs

  • Eleven Realtime v1 Mini (eleven_realtime_v1_mini): built for natural, low-cost conversation. ElevenLabs pitches it for AI assistants, tutors and interactive characters, and says it's enough for most use cases.
  • Eleven Realtime v1 Max (eleven_realtime_v1_max): reasons before it replies and handles multi-step tool calls, aimed at customer support, lead qualification and scheduling.

Both support 90+ languages and can speak with any of the 17,000+ voices in the ElevenLabs library. That's the part I like most: you're not stuck with a handful of stock agent voices, so a game character or a storytelling bot can keep the exact voice you already picked.

The numbers

ElevenLabs says Artificial Analysis evaluated both models on three public benchmarks. These are the figures from the launch post and the Eleven Realtime page:

  • Big Bench Audio (1,000 reasoning questions asked by voice): v1 Max 98.8%, v1 Mini 75.9%.
  • Time to first audio (average on Big Bench Audio): v1 Max 0.57 s, v1 Mini 0.62 s.
  • Full-Duplex-Bench (pauses, interruptions, turn-taking, backchannels): v1 Mini 95.8%, v1 Max 94.1%.
  • Speech Agent Arena (blind live conversations): v1 Max 998, v1 Mini 988.
  • Tool-use task success: v1 Max 87.1%, v1 Mini 60.3%.

So Mini is actually slightly better at the feel of a conversation, while Max pulls far ahead once the agent has to reason or call tools. Some coverage quotes faster times of around 470 ms; I'm sticking with the averages ElevenLabs published.

Pricing

The API pricing page lists both Realtime v1 Mini and Realtime v1 Max at $0.08 per minute plus LLM usage. Inside ElevenAgents you pay the standard ElevenAgents per-minute rate plus LLM tokens, and the model picker shows an estimated cost per minute. Through the Realtime API you pay per minute of connected audio plus tokens, with the same token prices on every plan.

Try It Yourself

No code: create an agent in ElevenAgents (or open one you already have) and pick a Realtime model from the model dropdown. You keep all the platform features: phone, web, WhatsApp and contact-center channels, workflows, a knowledge base, guardrails, testing and monitoring.

If you already use OpenAI's Realtime API: keep your code and change the endpoint and model name. This is the launch post's example, trimmed:

import { OpenAIRealtimeWebSocket, RealtimeAgent, RealtimeSession } from "@openai/agents-realtime";

const { value: clientSecret } = await fetch("/api/token").then((res) => res.json());

const agent = new RealtimeAgent({
  name: "ElevenLabs agent",
  instructions: "You are a helpful, friendly voice assistant. Keep replies concise.",
});

const transport = new OpenAIRealtimeWebSocket({
  url: "wss://api.elevenlabs.io/v1/convai/realtime",
});

const session = new RealtimeSession(agent, {
  transport,
  model: "eleven_realtime_v1_mini",
  config: { outputModalities: ["audio"], audio: { output: { voice: "brian" } } },
});

await session.connect({ apiKey: clientSecret });

Two catches from the docs: the Realtime API is WebSocket only (no WebRTC at launch), and it runs separately from ElevenAgents, so if you want channels or a knowledge base, build in ElevenAgents from the start.

A quick test script to say out loud, which pokes at interruptions and reasoning:

Hi! I want to book a table for four on Friday... actually, make that Saturday.
If a train leaves at 3:40 and the trip takes 95 minutes, when do I arrive?
Wait, stop. Say that again, but slower and in a pirate voice.

My tip: start with Mini for characters and companions, and switch to Max only when the agent has to look things up or follow a process.

Why It Matters for Makers

Until now, a talking character usually meant wiring together speech recognition, a turn detector, an LLM and a TTS voice, then tuning the gaps between them. Eleven Realtime collapses that into one model with ElevenLabs voices on the output end. That makes it a strong option for interactive NPCs, voice-driven story apps, live tutors and AI companions, especially if you've already designed or cloned a voice on ElevenLabs. The OpenAI-compatible endpoint also makes it cheap to A/B against other realtime models, like the Grok Voice 3 release we covered last week.

The trade-offs: it's API-only, the benchmark results are reported by ElevenLabs from Artificial Analysis evaluations, the LLM token cost sits on top of the $0.08 a minute, and WebRTC apps need to move to WebSocket first. As always, only use voices you have the rights to.

Conclusion

Eleven Realtime gives ElevenLabs two end-to-end speech-to-speech models: v1 Mini for natural, low-cost conversation and v1 Max with 98.8% on Big Bench Audio and 0.57 seconds to first audio, both with 17,000+ voices in 90+ languages. Pick one in ElevenAgents or point your OpenAI Realtime code at the new endpoint, and have a proper conversation with your favorite voice. Happy voice-making!

Original source:

https://elevenlabs.io/blog/realtime

Product page: Eleven Realtime · Pricing: ElevenLabs API pricing · Announcement:

—Anabel ♡