Evergreen Tools
|
← Back to AI Tools

Eleven Realtime: End-to-End Voice Agent Models

A class of speech-to-speech models launched 20 October 2026 where one model is a complete voice agent — recognition, turn-taking, the LLM and synthesis folded into a single model instead of a stack you assemble yourself

Official Source

Summarised from ElevenLabs' launch post of 20 October 2026; benchmarks were run by Artificial Analysis.

Features

  • ✓ Two tiers with a clear split: v1 Mini for low-cost natural conversation (assistants, teachers, characters) and v1 Max for complex tasks (support, lead qualification, multi-step workflows)
  • ✓ More than 90 languages and access to 17,000+ voices from the ElevenLabs library
  • ✓ Compatible with OpenAI's Realtime API spec, so an existing integration mostly means changing the endpoint and model name
  • ✓ Independent Artificial Analysis results: v1 Max answers 98.8% on Big Bench Audio and starts speaking in 0.57s on average, while v1 Mini scores 95.8% on Full-Duplex-Bench
  • ✓ Tool use is where they diverge: 87.1% completion for v1 Max versus 60.3% for v1 Mini, which is why the docs recommend Max for tool-heavy work

How to Use

  1. Decide the job first: chat-style assistants on Mini to save money, process-following and tool-calling agents on Max
  2. Start inside ElevenAgents and pick a Realtime model, which brings phone, web and WhatsApp channels plus knowledge base, guardrails and monitoring
  3. If you already run an OpenAI Realtime integration, point it at wss://api.elevenlabs.io/v1/convai/realtime and switch the model name to eleven_realtime_v1_mini or max
  4. Listen to it handle interruptions before you ship — full-duplex quality is an ear test, not a benchmark score

FAQ

What is Eleven Realtime?

A class of speech-to-speech models ElevenLabs launched on 20 October 2026, described in the announcement (https://elevenlabs.io/blog/realtime) as complete voice agents in a single model. Instead of choosing speech recognition, voice activity detection, turn-taking, an underlying LLM and text to speech separately, those decisions collapse into one model per use case.

How do Mini and Max differ?

ElevenLabs positions v1 Mini for natural conversation at low cost — AI assistants, teachers, interactive characters, sufficient for most cases — and v1 Max for complex tasks such as customer support, lead qualification and multi-step workflows that follow a process and reason through steps. The numbers back the split: 98.8% versus 75.9% on Big Bench Audio, and 87.1% versus 60.3% on tool-use completion.

What do the benchmarks and latency look like?

Artificial Analysis evaluated both models independently. On Big Bench Audio, 1,000 spoken reasoning questions, Max answered 98.8% and Mini 75.9%, with Max starting to speak after 0.57 seconds on average and Mini after 0.62. On Full-Duplex-Bench, which tests pauses, turn-taking, interruptions and backchannels, Mini scored 95.8% and Max 94.1%. In the blind Speech Agent Arena, Max scored 998 and Mini 988. Note that Mini leads on conversational dynamics while Max wins on reasoning and tools.

Do I have to rewrite an existing OpenAI Realtime integration?

Mostly not. The Realtime API runs over WebSocket and follows the OpenAI Realtime spec, so OpenAI's SDK works as is and migration means changing the endpoint and model name. Two limits matter: WebSocket only, so a WebRTC integration has to move first, and it runs separately from ElevenAgents, so if you need channels or a knowledge base, build on ElevenAgents instead.

When is it the wrong choice?

If your voice agent needs the surrounding operational stack — phone and WhatsApp channels, knowledge base, guardrails, testing and monitoring — go to ElevenAgents directly, since the Realtime API is the model layer only. And for very high-volume, very simple conversations where cost dominates, run the unit economics against a cheaper assembled pipeline before committing.