ElevenLabs has launched Eleven Realtime, two speech-to-speech models that bundle the components needed for a live voice agent. Instead of separately tuning speech recognition, turn detection, a language model, and text-to-speech, developers can choose a model built as a complete conversational system.
Realtime v1 Mini targets lower-cost assistants, teachers, and interactive characters. Realtime v1 Max is intended for customer support and multi-step workflows that require more reasoning and tool calls. Both support more than 90 languages and the company’s library of more than 17,000 voices.
ElevenLabs says an independent Artificial Analysis evaluation gave Max 98.8% accuracy on a 1,000-question spoken reasoning benchmark, while Mini reached 75.9%. Max completed 87.1% of tested tool-use tasks and Mini completed 60.3%. Those figures are benchmark results, not guarantees for production calls.
The models are available through ElevenAgents and a WebSocket Realtime API. The API follows OpenAI’s Realtime specification, so existing integrations can change the endpoint and model name, but WebRTC applications must first move to WebSocket.