ElevenLabs has released Eleven Realtime, a pair of speech-to-speech models for voice agents, and its pitch rests on one idea: the whole conversation loop should be tuned as a single system. The company, best known for synthetic voices, says speech recognition, turn-taking, the language model, and the voice are optimized together rather than stitched from separate parts.
The two versions split by job. v1 Mini is aimed at expressive, flowing talk, such as assistants, tutors, and interactive characters. v1 Max is aimed at work that needs precise answers, and ElevenLabs names customer support and lead qualification. That is a familiar trade between warmth and accuracy. The product page does not say how the two differ in size, speed, or price.
Inside ElevenAgents, the company’s platform for building agents, a team picks Mini or Max when it creates an agent, and transcription, turn-taking, and voice come preconfigured. According to ElevenLabs, the same agent can go out over phone, a web widget, WhatsApp, and contact center software without a separate integration for each. The platform also offers a knowledge base, guardrails, scripted test scenarios, and monitoring of live conversations.
The developer route matters more. Eleven Realtime is also available through an API that follows the shape of OpenAI’s Realtime API. ElevenLabs says a team swaps the endpoint and the model name and leaves the rest of its code alone. Messages flowing in both directions are formatted the way OpenAI’s are, which means existing event handlers, along with any logging and tooling a team has built, carry over. Compatibility like this removes the main cost of trying a rival: rewriting the plumbing. It also works in reverse, since a customer who can switch in cheaply can switch out just as cheaply.
Voice choice is the other selling point. ElevenLabs lists more than 17,500 voices, grouped into narration, conversational, social media, character, educational, advertisement, entertainment, and multilingual sets. Realtime conversation runs in over 90 languages. For regulated buyers, the company says it encrypts data both while it moves and while it is stored, and it lists compliance with SOC 2, HIPAA, and GDPR, plus EU data residency and zero-retention modes. A dashboard keeps a searchable log of each conversation, which the company pitches for audits.
Nothing on the page can be checked from outside. ElevenLabs gives no independent benchmarks, no response-time figures, no pricing, and no word on how the models handle interruptions or noisy phone lines, which are the places voice agents usually fail. Claims such as sounding natural are the company’s own description of its product.
Any team with a working prototype on OpenAI’s realtime interface can now point it at Mini and Max in an afternoon. The sensible move is to run your own recordings through both, measure delay and interruption handling yourself, and price the result before trusting the demo, because the vendor has supplied no numbers to do it for you.
Reported by ElevenLabs on its Eleven Realtime product page, as published this week (October 2026).
