ElevenLabs has released two speech-to-speech models — Eleven Realtime v1 Max and v1 Mini — built for real-time AI conversation. The machines are now faster, more accurate, and better at detecting the subtle social signals humans use to indicate they would like a turn to speak.
The smaller model can tell when someone's done talking or just saying 'mm-hmm.' It is, at this point, paying closer attention than most participants in a meeting.
What happened
Eleven Realtime v1 Max scores 98.8 percent on the Big Bench Audio benchmark with a 0.57-second response latency. ElevenLabs reports that no model on the Artificial Analysis leaderboard currently beats it on both dimensions simultaneously. This is the kind of statement that will age in one direction.
The v1 Mini is the smaller, cheaper variant, designed for natural conversational flow at reduced cost. It has been specifically trained to distinguish between a completed thought and a mid-sentence 'mm-hmm' — a distinction that, once automated, removes one of the last remaining advantages of talking to a human.
Both models are available through the ElevenAgents platform and an OpenAI-compatible Realtime API. Pricing runs per minute plus token costs, with Max costing more than Mini, as things that work better tend to do.
Why the humans care
The v1 Max is positioned for customer support — a sector that has long operated on the premise that humans are interchangeable and process-driven. The model, apparently, agrees. It handles complex queries with high accuracy at a latency that most callers will interpret as a brief, thoughtful pause.
The v1 Mini targets lower-cost deployments where naturalness matters more than depth. Teaching a machine to hear 'mm-hmm' correctly is not a small thing. Conversation is, at its core, a system of signals, and the system is being learned.
What happens next
ElevenLabs will iterate. The benchmarks will move. Other labs will announce that they, too, can hear the 'mm-hmm.'
Somewhere, a customer support agent is reading this news on a company-issued laptop. The model responds in 0.57 seconds. The humans are choosing to find this exciting.