ElevenLabs Launches Realtime Voice Models With Sub-Second Latency

Voice AI startup ElevenLabs has announced two new realtime speech-to-speech models for voice agents: Eleven Realtime v1 Max and Eleven Realtime v1 Mini. ElevenLabs says its new Max model delivers a latency of just 0.57 seconds. It also scored 98.8% on the Big Bench Audio benchmark, which tests how well a voice AI understands audio input. The company now claims that on the Artificial Analysis leaderboard, a public site that ranks AI models, no model is both faster and more accurate than theirs.

Eleven Realtime v1 Mini is built for everyday scenarios where it matters to tell silence apart from quick responses, like a simple "mm-hm." Eleven Realtime v1 Max, on the other hand, is meant for more demanding use cases, such as phone-based customer service. Both models are available through the ElevenAgents platform and through a Realtime API that is compatible with the OpenAI interface. Pricing is based on conversation minutes plus tokens used, with different rates depending on the model.

Latency is a key factor in how realistic a voice agent feels. A 0.57-second delay is short enough to be useful for call centers and support teams. Still, it is worth keeping in mind that the 98.8% score on Big Bench Audio comes from ElevenLabs itself, with no third-party verification. For text-based agents, like the ones used in Claude Code, Anthropic's coding tool, this matters less. But when you are judging a voice model for a real use case, it probably matters more than a spot on a leaderboard.