ElevenLabs launches Eleven Realtime, two speech-to-speech models with an OpenAI-compatible API

On 20 October 2026 ElevenLabs announced Eleven Realtime, which it describes as a new class of speech-to-speech models. Two models shipped: v1 Mini, positioned for natural conversation at lower cost, and v1 Max, positioned for complex multi-step tasks such as customer support, lead qualification and workflows that require reasoning through steps. Both support more than 90 languages and can use any voice in the ElevenLabs library, which the company puts at over 17,000. They are available through ElevenAgents and a Realtime API.

ElevenLabs says the Realtime API is compatible with OpenAI's Realtime API, so existing users can switch by changing the endpoint and model name; it connects over WebSocket only, so WebRTC integrations must migrate first. Artificial Analysis evaluated both models independently: on Big Bench Audio speech reasoning, v1 Max returned 98.8% correct answers against v1 Mini's 75.9%; average time to first audio was 0.57 seconds for Max and 0.62 seconds for Mini; Full-Duplex-Bench conversational dynamics came in at 94.1% for Max and 95.8% for Mini; Speech Agent Arena scores were 998 and 988; completed tool-use tasks were 87.1% and 60.3%. The announcement gives no pricing.

Why it matters for digital humans: Avatar vendors mostly do not build their own speech stack, so the voice layer's latency floor sets the ceiling on how natural a talking head feels. Two things here matter to practitioners. The OpenAI protocol compatibility turns the speech-to-speech layer into something closer to a swappable component, which weakens lock-in for anyone assembling an avatar from a rendering vendor plus a voice vendor. And the Mini/Max split is an explicit admission that reasoning quality and response latency still trade against each other: Mini edges Max on conversational dynamics while Max wins decisively on reasoning and tool use. An avatar doing scripted reception work and one doing agentic support are no longer the same purchase.