ElevenLabs launches Eleven Realtime: v1 Mini and v1 Max speech-to-speech models for voice agents, with an OpenAI-compatible Realtime API

Why it matters

【Key points】

ElevenLabs launched Eleven Realtime, a new class of speech-to-speech models for voice agents, starting with v1 Mini and v1 Max. Speech recognition, turn-taking, the LLM and speech synthesis are tuned as one system, so one model choice gives a complete voice agent.

1. Two models: v1 Mini is the default, with the lowest latency and cost, for assistants, tutors and characters. v1 Max reasons before answering and handles multi-step tool calls, for support and scheduling; it costs more.

2. ElevenLabs' own figures: Mini, 620 ms to first audio, 95.8% on Full-Duplex-Bench, Speech Agent Arena Elo 988. Max, 570 ms to first audio, 98.8% on Big Bench Audio, 87.1% task success. Both cover 90+ languages and 17,500+ voices.

3. Access: set an agent's voice mode to Realtime in ElevenAgents, or use the Realtime API, which is compatible with OpenAI's Realtime API. You change the endpoint and model name (e.g. eleven_realtime_v1_mini) and keep your code.

4. Limits: the rollout is gradual. Custom LLMs and Amazon Bedrock are not supported. Zero Retention Mode and HIPAA only work with Max, in the US and EU, and Max is not offered for India or Singapore data residency. Billing is the per-minute ElevenAgents rate plus model tokens.

【Verifiable】

Product page: https://elevenlabs.io/realtime

Realtime models docs: https://elevenlabs.io/docs/eleven-agents/customization/realtime-models

Models overview: https://elevenlabs.io/docs/overview/models

【Our read】

1. Voice agents are moving from chained ASR + LLM + TTS pipelines to single models, as OpenAI and SpaceXAI have done. Matching OpenAI's API is a clear bid to make switching cheap.

2. All numbers are self-reported, and the arena and task-success methods are not detailed. There is no blog post or report yet; the product page and docs are the only primary sources.

3. ElevenAgents users skip component tuning; teams needing a specific LLM or strict compliance should stay in custom mode.