ElevenLabs launched Eleven Realtime, end-to-end speech-to-speech models meant to replace separately wired ASR + VAD + turn-taking + LLM + TTS stacks. Two SKUs: Realtime v1 Mini for low-cost natural conversation (assistants, tutors, characters) and Realtime v1 Max for complex support / multi-step workflows with tool use. Both claim 90+ languages and access to ElevenLabs' 17,000+ voice library. Artificial Analysis (vendor-cited independent) scores: Big Bench Audio accuracy 98.8% Max / 75.9% Mini; time-to-first-speech ~0.57s Max / ~0.62s Mini; Full-Duplex-Bench 94.1–95.8%; Speech Agent Arena Elo ~998 Max / ~988 Mini; tool-use completion 87.1% Max / 60.3% Mini. Delivery via ElevenAgents UI or an OpenAI Realtime API–compatible WebSocket (wss://api.elevenlabs.io/v1/convai/realtime, models eleven_realtime_v1_mini / eleven_realtime_v1_max). Vendor-confirmed; desk did not independently reproduce latency/accuracy.
Clearest same-window live/voice model product ship — OpenAI-compatible realtime endpoint lowers migration cost for teams already on Realtime APIs.
More ▾ Less ▴
ElevenLabs published Introducing Eleven Realtime on Oct 20, 2026 (on-page Published Oct 20; editor spot-check ~06:30 HKT Wed 21). The post frames Realtime as speech-to-speech models that collapse ASR/VAD/turn-taking/LLM/TTS into one stack, with Mini for natural low-cost conversation and Max for complex support and multi-step tool workflows. Both support 90+ languages and the 17,000+ voice library. Vendor-cited Artificial Analysis numbers: Big Bench Audio 98.8% Max / 75.9% Mini; first speech ~0.57s Max / ~0.62s Mini; Full-Duplex-Bench 94.1–95.8%; Speech Agent Arena Elo ~998 / ~988; tool-use 87.1% / 60.3%. Delivery is ElevenAgents UI or OpenAI Realtime–compatible WebSocket at wss://api.elevenlabs.io/v1/convai/realtime (models eleven_realtime_v1_mini / eleven_realtime_v1_max).
- Pricing page lock for connected-minute and token rates
- Enterprise CCaaS channel uptake
- Whether Max tool-use holds on third-party evals
Vendor-confirmed ElevenLabs primary (Oct 20). Artificial Analysis scores are vendor-cited; editor did not independently reproduce latency or accuracy. Pricing dollars not yet locked from docs for Wire voice.