ElevenLabs introduces Eleven Realtime, a family of speech-to-speech models for voice agents

ElevenLabs announced Eleven Realtime, a family of speech-to-speech models for voice agents, on October 20. The company says each model tunes speech recognition, turn-taking, the language model and text-to-speech together as one system. Developers no longer pick and configure those parts separately, and existing prompts, voices, tools and knowledge bases keep working. There are two versions. v1 Mini, the default and cheapest option, targets assistants, tutors and interactive characters. v1 Max reasons before it answers and is recommended for customer support and multi-step tool calls.

In results published by the company and described as independently measured by Artificial Analysis, time to first audio is 620 milliseconds for Mini and 570 for Max. Max scores 98.8% on Big Bench Audio, a spoken reasoning test, and 87.1% on task success. Mini scores 95.8% on Full-Duplex-Bench and 988 Elo on the Speech Agent Arena. In its post on X, ElevenLabs said voice agents usually force a choice between quick replies and precise answers, and that the new models offer both. The models converse in more than 90 languages.

The models are available in the ElevenAgents platform and through the Realtime API, and the changelog says they are rolling out gradually across workspaces. The Realtime API is compatible with OpenAI’s Realtime API, and ElevenLabs says migrating takes only a new endpoint and model name. At launch it supports WebSocket only, not WebRTC. Pricing combines per-minute fees with per-token LLM costs, but the launch page lists no figures. The same update opened a Realtime Translation API for live speech translation to enterprise workspaces.