ElevenLabs Launches Eleven Realtime Speech-to-Speech Models for Voice Agents

ElevenLabs on Oct. 20, 2026 launched Eleven Realtime, a new family of speech-to-speech models that bundle speech recognition, turn-taking, a language model and voice output into a single system for building AI voice agents. The company released two versions, v1 Mini and v1 Max, and made them available the same day in its ElevenAgents platform and through its Realtime API.

Image: A pair of studio condenser microphones, used for illustration. Photo: Tim Walker, via Wikimedia Commons (CC BY 2.0).

What Eleven Realtime is

Building a voice agent has usually meant stitching together several separate pieces: one model to transcribe speech, another to detect when a person has stopped talking, a large language model to decide what to say, and a text-to-speech model to say it. In a launch post written by Min Kim, ElevenLabs said that getting all of those parts to work smoothly together “is a project in itself.”

Eleven Realtime is the company’s answer. Each model is pitched as a complete voice agent, with those components tuned together end to end. Both models support more than 90 languages and can use any of the more than 17,000 voices in the ElevenLabs voice library, the company said.

The French tech outlet Journal du Net described the release as a move away from the multi-model setup ElevenLabs customers previously used for voice AI, toward one model that handles the whole conversation.

Eleven Realtime v1 Mini vs. v1 Max

ElevenLabs is positioning the two models for different jobs:

• v1 Mini is the lower-cost option, built for natural, expressive conversation. The company recommends it for AI assistants, tutors and interactive characters, and says it will be enough for most uses.
• v1 Max adds more reasoning for complex tasks such as customer support, lead qualification, scheduling and multi-step workflows where the agent has to follow a process and call outside tools.

ElevenLabs’ developer documentation lists the model IDs as eleven_realtime_v1_mini and eleven_realtime_v1_max, and notes that Max costs more.

The benchmark numbers ElevenLabs is citing

The company published a set of test results alongside the launch. These are ElevenLabs’ own figures, not independent tests:

• Reasoning by voice: On Big Bench Audio, which asks 1,000 reasoning questions out loud and scores the spoken answers, v1 Max answered 98.8% correctly and v1 Mini 75.9%.
• Speed: v1 Max starts speaking in about 0.57 seconds on average, and v1 Mini in about 0.62 seconds.
• Conversation flow: On Full-Duplex-Bench, which tests pauses, interruptions and turn-taking, v1 Mini scored 95.8% and v1 Max 94.1%.
• Tool use: v1 Max completed 87.1% of tool-use tasks, compared with 60.3% for v1 Mini, which is why the company recommends Max for heavy tool calling.
• Blind comparisons: In the Speech Agent Arena, where people compare two models in live conversations without knowing which is which, v1 Max scored 998 and v1 Mini 988.

ElevenLabs argued that v1 Max breaks the usual trade-off in voice AI, where the most accurate models tend to be the slowest to answer.

Built to be a drop-in for OpenAI’s Realtime API

One of the more practical details is compatibility. ElevenLabs says its Realtime API speaks the same wire protocol as OpenAI’s Realtime API, with the same client and server events over a WebSocket connection. Developers who already built an app on OpenAI’s Realtime SDK can point it at ElevenLabs by changing the endpoint and model name, without switching SDKs, according to the company’s guide.

Journal du Net reported that WebRTC connections are not supported at launch. Inside ElevenAgents, picking a Realtime model is a single dropdown choice, and agents built that way can be deployed by phone, on websites and through other channels the platform supports. ElevenLabs said existing agents will not change unless their owners switch them over.

Why it matters

Voice is becoming one of the busiest fronts in AI, as companies try to replace phone trees and scripted chatbots with agents that can hold a real conversation. A single model that handles the full loop could make those agents cheaper and faster to build, if the performance holds up outside the company’s own tests.

The launch lands in a crowded week for AI news. The Boston Report has covered Disney’s $20 million Character.AI license deal and the bankruptcy fight around it, Tenstorrent’s $1.5 billion raise for AI chips, and Boston’s own AI startup Graylark and its $10.7 million seed round.

Sources

• ElevenLabs: Introducing Eleven Realtime, a simpler way to build natural voice agents
• ElevenLabs: Eleven Realtime product page
• ElevenLabs Docs: OpenAI Realtime API compatibility
• Journal du Net: ElevenLabs lance Eleven Realtime
• Wikimedia Commons: Behringer C-2 condenser microphone pair (CC BY 2.0)