ElevenLabs ships Eleven Realtime, a 0.57-second voice model
A new speech-to-speech model family in Mini and Max tiers, with an API wire-compatible with OpenAI Realtime, aimed squarely at voice-agent builders.
ElevenLabs released Eleven Realtime on 20 October, a family of speech-to-speech models built for voice agents rather than general chat. There are two tiers: v1 Mini, tuned for low-latency conversation, and v1 Max, tuned for tasks that need reasoning and tool calls.
ElevenLabs reports v1 Max starts speaking after 0.57 seconds on average, against 0.62 seconds for v1 Mini. On its Big Bench Audio benchmark, Max scores 98.8% and Mini 75.9%; on tool-use tasks, 87.1% against 60.3%. Both models support more than 90 languages and any of the 17,000-plus voices in ElevenLabs’ library. The Realtime API runs over WebSocket and is wire-compatible with OpenAI’s Realtime API, and the models are also reachable through ElevenLabs’ managed ElevenAgents platform for phone, web, WhatsApp and contact-centre deployments.
Why it matters: OpenAI-compatible wire format means swapping a voice agent’s backing model is closer to a config change than a rewrite. The Mini and Max split also makes the latency-versus-accuracy trade explicit, so you can route simple turns to the cheap, fast model and only pay for reasoning when a call actually needs it.
The caveat: every number above comes from ElevenLabs’ own announcement, not an independent benchmark run.