Lifestyle
ElevenLabs Launches Eleven Realtime: Building Voice AI Without Assembling Every Component Yourself
ElevenLabs has announced Eleven Realtime v1 Mini and v1 Max, saying developers can build AI that talks by voice with a single model, without choosing and tuning many components. This may cut work for teams building voice support or assistants, but the API supports WebSocket connections only.
About 5 min read
What happened
In a blog post published on October 20, 2026, ElevenLabs announced Eleven Realtime, which it describes as a new class of Speech to Speech model: speech goes in, and speech comes out.
According to ElevenLabs, to build a voice agent that sounds natural, developers must make separate choices on several components, including speech recognition, voice activity detection (deciding whether the other party is speaking), turn-taking, the underlying large language model and speech synthesis, and getting them to work well together takes further testing and tuning. ElevenLabs says Eleven Realtime is meant to save that work.
ElevenLabs says two models launch first: v1 Mini focuses on low-cost natural conversation, suited to AI assistants, teachers and interactive characters; v1 Max is for complex tasks, suited to customer support, lead qualification (sorting out which potential customers to pursue) and multi-step workflows. The company says both models support more than 90 languages and can use more than 17,000 voices from its voice library.
How it works and how the two models compare
ElevenLabs says Eleven Realtime is built on an end-to-end tuned cascaded architecture, meaning multiple components process the conversation in sequence and are then tuned as a whole. What users face is only audio in, audio out and a single model ID. The company notes that every step produces text, so one can see what the agent heard and decided to say; this text also lets ElevenAgents (ElevenLabs' product for building and running voice agents) apply guardrails (safety mechanisms limiting agent behavior) and PII redaction (hiding personal data) to Realtime agents.
| Item | v1 Mini | v1 Max |
|---|---|---|
| Positioning (ElevenLabs' wording) | Low-cost natural conversation | Complex tasks, multi-step workflows |
| Big Bench Audio accuracy (answering 1,000 reasoning questions by voice) | 75.9% | 98.8% |
| Average time to start speaking | 0.62 s | 0.57 s |
| Full-Duplex-Bench (tests pauses, turn-taking, interruptions and backchannels, the brief acknowledgements a listener gives while someone else talks) | 95.8% | 94.1% |
| Speech Agent Arena score (real people compare two models in live conversation without knowing which is which) | 988 | 998 |
| Tool-use task completion rate (tasks where the AI has to call on other software tools) | 60.3% | 87.1% |
ElevenLabs also says that on the Artificial Analysis leaderboard, no model is both faster and more accurate than v1 Max, and, given the difference in tool-use performance, it recommends Max for scenarios that call on tools heavily.
What it means for developers
ElevenLabs says its Realtime API is compatible with OpenAI's Realtime API, so developers trying it only need to change the endpoint (the URL the code connects to) and the model name in their code. The company also says the cascaded architecture makes the text at each step inspectable, which helps tell whether a problem is a misheard word or a faulty decision.
How to get started and limitations
ElevenLabs offers two routes:
- ElevenAgents (recommended by the company): after creating or opening an agent, pick a Realtime model from the dropdown. ElevenAgents provides phone, web, WhatsApp and CCaaS (cloud contact-center service) channels, plus workflows, knowledge bases, guardrails, testing and monitoring.
- Realtime API: developers who have already integrated the OpenAI Realtime API can reuse their code and change where it connects. ElevenLabs says its migration guide covers the steps and differences.
On limitations, ElevenLabs says the Realtime API connects over WebSocket only (one way of making a network connection); integrations that use WebRTC (another real-time connection technology) must first switch to WebSocket. It also runs separately from ElevenAgents, so developers who need platform features such as channels or knowledge bases should build on ElevenAgents from the start.
Frequently asked questions
What is Eleven Realtime?
According to ElevenLabs, it is a class of speech-to-speech models in which each model is a complete voice agent. The first two are v1 Mini and v1 Max.
What is the difference between Mini and Max?
ElevenLabs says Mini suits low-cost natural conversation and Max suits complex tasks. In the tool-use task completion rates the company published, Max scores 87.1% and Mini 60.3%.
Who published these benchmark figures?
The figures were published by ElevenLabs in its own article, and the company says the evaluations were run by Artificial Analysis. Treat them as vendor-published information.
Can code written for the OpenAI Realtime API be used directly?
ElevenLabs says it is compatible and only the endpoint and model name need changing; however, it supports WebSocket only, so integrations using WebRTC must first switch to WebSocket.
Browse the latest news in this topic
Lifestyle
AI Data Center Backlash in Maine Becomes a Senate Race Flashpoint, The Verge Reports
The Verge reports that many Maine residents worry AI data centers raise power bills, drain resources and bring few lasting jobs. That backlash has become a Senate race issue, and the result will help decide which party controls the Senate and shape US data center and AI policy.
Lifestyle
Wikipedia's Operator Tells Courts: AI Training Can Be Fair Use, but Chatbots Repeating Articles Should Need a License
According to The Verge, the Wikimedia Foundation, which runs Wikipedia, has filed a court brief taking neither side in the lawsuits by publishers including The New York Times against OpenAI and Microsoft. It argues for treating AI training and AI output differently, and warns that an oversimplified ruling could cost Wikipedia editors useful AI tools or endanger the open web.
Lifestyle
No Belief in a God Required? A Humanist Wins an Exemption From Using AI at Work
According to TechCrunch, an administrator at the University of Kentucky has obtained a religious exemption from using AI at work, based on humanist beliefs. This piece covers what happened, the university's position and what it means for ordinary workers.
Lifestyle
Amazon CEO: AWS Could Become a $1 Trillion Business, but First Comes About $220 Billion in Spending This Year
According to TechCrunch, Amazon CEO Andy Jassy said on the earnings call that AWS, Amazon's cloud business, could reach $1 trillion in annual revenue within five to ten years. Heavy AI infrastructure investment comes first: cash capital expenditures this year are expected at about $220 billion, and investors are concerned about the spending.