ElevenLabs introduces Eleven Realtime voice models for conversations and complex tasks

Mini is aimed at assistants, tutors and interactive characters. ElevenLabs describes it as handling interruptions and quick exchanges. Max is intended for work such as customer support and lead qualification, including multi-step requests that call tools to look things up or take action.

What the speed figures measure

ElevenLabs lists time to first audio at 620 milliseconds for Mini and 570 milliseconds for Max. Its FAQ says those figures are averages measured on Big Bench Audio. Neither listed average is less than half a second, despite the company's broader launch claim that Eleven Realtime starts talking in half a second.

The company also reports a 98.8% Big Bench Audio score and an 87.1% task success rate for Max. Its FAQ says Artificial Analysis evaluated the models independently.

ElevenAgents and the API are separate routes

At launch, ElevenLabs said both models were available in ElevenAgents and Eleven Realtime was also available through its Realtime API. In ElevenAgents, builders can select Mini or Max and use phone, web-widget, WhatsApp and contact-center channels, plus a knowledge base, guardrails, testing and monitoring. The company's FAQ says ElevenAgents and the Realtime API run separately; an API-only integration does not include those platform features.

ElevenLabs calls the Realtime API OpenAI-compatible. Its product page says existing OpenAI Realtime API users can change the endpoint and model name, but also points to a migration guide for differences. The API connects over WebSocket and did not support WebRTC at launch, so an existing WebRTC integration would need to move to WebSocket first.

Many accounts welcomed Eleven Realtime for combining fast 500ms responses with accurate reasoning to advance voice agents, while some objected to its higher pricing than Grok or GPT and questioned the timing.

Based on 32 sentiment-bearing replies from 32 accounts across 1 conversation.

Many accounts welcomed Eleven Realtime for combining fast 500ms responses with accurate reasoning to advance voice agents, while some objected to its higher pricing than Grok or GPT and questioned the timing.

Based on 32 sentiment-bearing replies from 32 accounts across 1 conversation.