
ElevenLabs Launches Eleven Realtime, a New Class of Speech-to-Speech Models for Voice Agents
ElevenLabs has released two speech-to-speech models, Eleven Realtime v1 Max and v1 Mini, aimed at natural AI voice conversations. The Max variant scores 98.8 percent on the Big Bench Audio benchmark with a 0.57-second response time.
The voice-agent race has a new entry point. On Tuesday, October 20, ElevenLabs launched Eleven Realtime, what it calls a new class of speech-to-speech models built for natural, low-latency AI conversations.
Two models ship together. Eleven Realtime v1 Max is the flagship: it scores 98.8 percent on the Big Bench Audio benchmark with a 0.57-second response time. ElevenLabs says no model on the Artificial Analysis speech-to-speech leaderboard is both faster and more accurate right now. The smaller v1 Mini is built for natural everyday conversations at lower cost — and, tellingly, it can tell when someone has finished talking versus just saying “mm-hmm.”
Built for agents, not just audio
The positioning matters as much as the numbers. V1 Max targets complex tasks like customer support, where accuracy and reliability decide whether a deployment survives contact with real callers. V1 Mini targets the casual conversational layer — the chit-chat, intake, and triage calls where cost per minute is the gating factor and a missed turn is a minor annoyance rather than a failed transaction.
This is the part of voice AI that has been hardest to productize. Text-to-speech quality stopped being the bottleneck some time ago; the friction now sits in the turn-taking layer. A voice agent that talks over a caller, waits too long, or mistakes a backchannel “mm-hmm” for a finished thought feels broken no matter how good its voice sounds. ElevenLabs is explicitly selling Mini on this distinction — a model that knows when you are done versus when you are merely listening. If that capability holds up in production, it solves one of the most consistent complaints about deployed voice agents.
OpenAI-compatible, priced per minute
Both models are available through the ElevenAgents platform and an OpenAI-compatible Realtime API, which lowers the integration barrier for developers already building on OpenAI’s stack. Pricing is per minute plus token costs, with v1 Max priced above v1 Mini; full details are on ElevenLabs’ pricing page.
The per-minute-plus-tokens structure is worth noting. Speech-to-speech models consume tokens on both the audio and reasoning sides, and per-minute pricing keeps the unit economics legible for high-volume deployments like support lines — the exact segment Max is aimed at.
A crowded week for voice
The launch lands days after Mistral AI released its own voice stack — Voxtral 3, a family of speech-to-text, text-to-speech, and voice-agent models with a LiveKit-based SDK. Voxtral Small Transcribe 3 covers more than 50 languages, and the TTS side supports over 35 languages and 50 dialects with emotional adaptation, with transcription priced from $0.0035 per minute. Mistral’s pitch centers on control: run it on your own infrastructure or through EU-hosted APIs to keep data in your hands.
The two releases frame voice agents differently. Mistral leans on open infrastructure and data sovereignty; ElevenLabs competes on raw conversational quality and turn-taking naturalness. One sells where the compute runs, the other sells how the conversation feels.
What to watch
As voice agents move from demos to real deployments, the differentiator is increasingly not whether the model can speak, but whether it can listen: knowing when to wait, when to interrupt, and when a “mm-hmm” is not an answer. ElevenLabs is betting that the hardest part of the problem — and the part worth paying for — is the turn. The Big Bench Audio score and the 0.57-second latency are table stakes; the turn-taking is the product.

