ElevenLabs, the voice AI startup best known for eerily human text-to-speech, just went fully native. On Tuesday, the company announced Eleven Realtime v1 — a new family of true speech-to-speech models built from the ground up as conversational voice agents, not stitched-together chains.
Until now, building a voice agent on ElevenLabs meant chaining models: speech-to-text, then an LLM, then text-to-speech. That era is officially over. Realtime v1 listens and responds in one model, and it handles pauses, interruptions, and turn-taking like a real conversation.
Two models, one goal: lose the chain
The family ships in two flavors. Eleven Realtime v1 Mini is built for simple natural-language conversations where latency is the priority — ElevenLabs says it's enough for most use cases. Eleven Realtime v1 Max is the heavyweight: reasoning and accuracy first, aimed at customer service, agent workflows, and anything that needs multi-step reasoning with tools. Both run across 90 languages and can use the 17,000 voices in the ElevenLabs library.
The numbers are legitimately fast
On independent benchmarks from Artificial Analysis, Max puts up numbers that top the current voice charts. On Big Bench Audio's voice-reasoning test it scores 98.8% — ahead of Grok Voice 3.0 at 98.5% and Gemini 3.8 Live Extended Thinking at 97.7% (Mini scores 75.9%). For conversational dynamics — how much it feels like talking to a person — Mini scores 95.8% and Max 94.1%, on par with OpenAI's GPT-Realtime 2.1. Where the gap really opens up is tool calls: Max succeeds 87.1% of the time versus 60.3% for Mini. In the blind Speech Agent Arena, Max scores 998 points to Mini's 988.
Then there's speed. Max responds in 470 milliseconds, Mini in 500 — versus 810ms for Grok Voice 3.0 and 1.14 seconds for GPT-Realtime-2 (High). On Artificial Analysis' chart, no model is both more accurate and faster than Max. Sub-half-second responses cross a psychological threshold: callers stop noticing they're talking to an AI.
It plugs straight into OpenAI's plumbing
Here's the chess move. The Realtime API is fully compatible with OpenAI's Realtime API — developers migrating a voice-agent workflow from OpenAI just change the endpoint and the model name, and the OpenAI SDK works as-is. The models are available from today inside ElevenLabs and ElevenAgents (deployment by phone, web, WhatsApp, or contact-center platforms), plus the company's own Realtime API over WebSocket — WebRTC is not supported at launch.
On pricing, ElevenLabs says ElevenAgents follows its standard per-minute pricing, and the Realtime API charges per minute of connected audio plus tokens — though the per-token price hasn't been announced yet.
Why this matters
This launch is really about enterprise. Large organizations now account for 55% of ElevenLabs' revenue, up from 40% just twelve months ago, with customers like Deutsche Telekom, Revolut, and Klarna. The production use cases are already clear: customer service, tech support, tutoring, sales advice, lead qualification. Voice agents aren't a demo anymore — they're infrastructure, and the vendor that makes them feel instantaneous wins.
For context, this is the same company whose text-to-speech
we compared against Murf earlier this month — but this is a different beast entirely, a native voice-to-voice model rather than a pipeline. And it's landing in the same wave as Google's director-grade AI voices in
Gemini 3.8 Flash TTS.
The take
The voice-agent race has moved on from "who sounds human" — everyone sounds human now. It's about speed plus reliability. A model that answers in 470ms and successfully calls tools 87% of the time is a support agent that never sleeps, never needs a lunch break, and never puts you on hold. The phone call just became software's newest battleground.
Comments
Post a Comment