ElevenLabs has launched Eleven Realtime, which it calls “a new class of Speech to Speech models” designed to replace the patchwork of components behind most voice agents. In a blog post on 20 October 2026, the company introduced two models: “Eleven Realtime v1 Mini for natural conversations at low cost” and “Eleven Realtime v1 Max for complex tasks”.
The pitch is simplicity. Building a natural-sounding voice agent today usually means stitching together speech recognition, voice activity detection, turn-taking logic, a language model and text to speech. ElevenLabs says each Realtime model is “a complete voice agent”, and that its API is compatible with OpenAI’s Realtime API, so existing code can switch with a few changes.
In brief
- ElevenLabs has launched two speech-to-speech models, Eleven Realtime v1 Mini and v1 Max, each a complete voice agent in one model.
- Both “support more than 90 languages” and can use “any of the 17,000+ voices” in the ElevenLabs library, the company says.
- ElevenLabs reports benchmark results from Artificial Analysis, including 98.8% on Big Bench Audio for v1 Max; no pricing was given.
Two models, two jobs
ElevenLabs positions Mini as “the best choice for AI assistants, teachers, and interactive characters”, adding that it “will be sufficient for most use cases.” Max is described as “ideal for customer support, lead qualification, and multi-step workflows, where the agent has to follow a process and reason through steps.”
The company says the hard part of voice agents lies in small judgement calls, such as how long to wait after a pause, or how to recover when a user interrupts. In ElevenAgents, its agent platform, these remain settings developers can adjust. Users can select a Realtime model from a dropdown and use ElevenAgents’ phone, web, WhatsApp and contact-centre channels, along with workflows, a knowledge base, guardrails, testing and monitoring.
For developers already using OpenAI’s Realtime API, ElevenLabs says trying Eleven Realtime “simply means changing the endpoint and model name”. The API connects over WebSocket only, so WebRTC integrations need to move to WebSocket first, and it runs separately from ElevenAgents.
How ElevenLabs says they perform
ElevenLabs says Artificial Analysis evaluated both models independently across three public benchmarks: Big Bench Audio for speech reasoning, Full-Duplex-Bench for conversational dynamics, and the Speech Agent Arena for human preference and task success. All figures below are as ElevenLabs reports them.
| Benchmark (as reported by ElevenLabs) | v1 Max | v1 Mini |
|---|---|---|
| Big Bench Audio (speech reasoning) | 98.8% | 75.9% |
| Average time to start speaking | 0.57 seconds | 0.62 seconds |
| Full-Duplex-Bench (turn-taking, interruptions) | 94.1% | 95.8% |
| Speech Agent Arena (blind human comparison) | 998 | 988 |
| Tool-use tasks completed | 87.1% | 60.3% |
Big Bench Audio asks 1,000 reasoning questions by voice and scores the spoken answer. Mini edges Max on Full-Duplex-Bench, which tests pauses, turn-taking, interruptions and backchannels. The tool-use gap is wide, which ElevenLabs says is “why we recommend Max for use cases with heavy tool calling.”
ElevenLabs also argues that Max breaks the usual trade-off between accuracy and speed. “The models that score highest on Big Bench Audio take the longest to answer, and the fastest ones give up a lot of accuracy,” the company wrote. “v1 Max excels in both of these metrics.” That is ElevenLabs’ reading of the results.
What wasn’t announced
The blog post gives no pricing for either model beyond describing Mini as “low cost”. It also does not name the language models underneath or give comparative figures for rival systems in text, although its charts show other models.
Our read
Folding the voice-agent stack into one model is where the market is heading, and OpenAI compatibility lowers the cost of switching. ElevenLabs’ voice library and multilingual support are real advantages. But the benchmark numbers come from the company’s own announcement, and Mini’s tool-use score suggests the cheaper model still has clear limits. Pricing will decide how far it gets.
Sources
- ElevenLabs: “Introducing Eleven Realtime: a simpler way to build natural voice agents” (Min Kim, 20 October 2026).
Related stories
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
