ElevenLabs launches two speech-to-speech models that each run a full voice agent

ElevenLabs has launched Eleven Realtime, a new class of speech-to-speech models in two sizes: v1 Mini and v1 Max. Each model is a complete voice agent. Normally you pick and tune speech recognition, voice activity detection, turn-taking, a language model and text to speech separately. Here one model replaces that stack. Both support more than 90 languages and any of the 17,000-plus voices in the ElevenLabs library.

The API is compatible with OpenAI's Realtime API, so existing code needs only a new endpoint and model name. There is one catch: it connects over WebSocket only, so a WebRTC integration has to move to WebSocket first. It also runs separately from the ElevenAgents platform, which handles phone, web, WhatsApp and contact-center channels, workflows, knowledge base, guardrails, testing and monitoring.

Mini is pitched at low-cost natural conversation, such as assistants, teachers and interactive characters. Max targets complex work like customer support, lead qualification and multi-step workflows. In an independent evaluation, Max scored 98.8% on 1,000 spoken reasoning questions against 75.9% for Mini. Time to first audio averaged 0.57 seconds for Max and 0.62 for Mini. On turn-taking and interruption handling, Mini scored 95.8% and Max 94.1%. In blind human preference, Max scored 998 and Mini 988.

The biggest gap is tool use. Max completed 87.1% of tool-use tasks, while Mini completed 60.3%. ElevenLabs therefore recommends Max for heavy tool calling.

Our take

The hard part of a voice agent was never the voice, it was wiring several parts together and tuning them. If that collapses into one model name in a config file, building voice agents gets much easier, and the competition shifts to price and reliability.