October 20, 2026 — ElevenLabs has released Eleven Realtime, a pair of speech-to-speech models — v1 Max and v1 Mini — built to hold live voice conversations without stitching together separate speech recognition, language model, and text-to-speech systems. Why it matters: ElevenLabs is moving from selling voices to selling the whole voice agent, in the same week its rivals are doing exactly the same.
What happened
Until now, building a voice agent on ElevenLabs meant assembling several models: one to transcribe speech, a language model to reason, and one to speak the reply. Eleven Realtime folds those pieces into a single speech-to-speech system, with recognition, turn-taking, reasoning, and voice tuned together.
The family has two members. Eleven Realtime v1 Max targets demanding work such as customer support, where accuracy and tool use matter most. Eleven Realtime v1 Mini is the cheaper option for natural, everyday conversation, and it can tell the difference between a speaker who has finished talking and one who is only saying “mm-hmm.”
The details
ElevenLabs says v1 Max scores 98.8 percent on the Big Bench Audio benchmark with a 0.57-second response time, and that no model on the Artificial Analysis speech-to-speech leaderboard is currently both faster and more accurate. Those are the company’s own launch claims; independent testing has not yet been published.
Reporting on the launch adds useful texture: the models support up to 90 languages and can draw on the more than 17,000 voices in the ElevenLabs library. On tool calling — the skill a support agent needs to actually look things up mid-call — Max reportedly succeeds 87.1 percent of the time, against 60.3 percent for Mini, which is the clearest signal of where the price difference goes.
Both models are available now through the ElevenAgents platform and through a Realtime API that is compatible with OpenAI’s realtime format, which lowers the switching cost for teams already built on OpenAI’s interface. Pricing is charged per minute of audio plus token usage, with Max priced above Mini; exact rates are on ElevenLabs’ pricing page.
Why it matters
Voice is having its platform moment: Mistral launched its Voxtral 3 agent stack the same day, and every major lab now wants to own the full conversation loop rather than one component of it. ElevenLabs’ advantage is its voice library and brand with creators; its challenge is that a speech-to-speech model has to reason as well as it speaks, which is newer territory for the company. For businesses buying voice agents, the practical takeaway is simpler: end-to-end models from the voice specialists are now a real alternative to assembling your own pipeline, and the OpenAI-compatible API means you can test that claim without rebuilding your stack.
Frequently asked questions
What is Eleven Realtime?
Eleven Realtime is ElevenLabs’ speech-to-speech model family for live voice agents. It combines speech recognition, reasoning, turn-taking, and voice output in one system, instead of chaining separate transcription, language model, and text-to-speech components.
What is the difference between v1 Max and v1 Mini?
Max is built for accuracy-heavy work like customer support — 98.8 percent on Big Bench Audio per ElevenLabs, and much stronger tool calling. Mini is cheaper and tuned for natural everyday conversation, including detecting whether a speaker has finished or is just saying “mm-hmm.”
How do developers access Eleven Realtime?
Through the ElevenAgents platform or an OpenAI-compatible Realtime API. Pricing is per minute of audio plus token costs, with Max costing more than Mini.
How many languages and voices does it support?
Launch reporting puts support at up to 90 languages, with access to the 17,000-plus voices in ElevenLabs’ library.
Sources: The Decoder, Journal du Net, ElevenLabs