ElevenLabs ships Eleven Realtime agents answering in 0.57 seconds

ElevenLabs shipped Eleven Realtime on October 20, two speech-to-speech models that each work as a complete voice agent.

The models arrived in an ElevenLabs blog post written by Min Kim and published on October 20, 2026, and they are live in ElevenAgents and in a Realtime API. Each combines speech recognition, voice activity detection, turn-taking, a language model and text to speech, tuned together, so a developer picks a model instead of wiring five.

Eleven Realtime v1 Mini is the lower-cost model, built for AI assistants, educational tutors and interactive characters. Eleven Realtime v1 Max adds reasoning and multi-step tool calls for customer support and lead qualification. Both support more than 90 languages and the library's 17,000+ voices.

Artificial Analysis evaluated both on three public benchmarks, per the launch post. On Big Bench Audio, 1,000 reasoning questions asked by voice, v1 Max answered 98.8% correctly and v1 Mini 75.9%. v1 Max starts speaking in 0.57 seconds on average and v1 Mini in 0.62 seconds. On Full-Duplex-Bench, v1 Mini reached 95.8% against 94.1% for v1 Max, and on tool-use tasks v1 Max completed 87.1% against 60.3%.

The evaluator's own Speech to Speech leaderboard lists both models, with v1 Max at 99% on speech reasoning, 87.1% task success and 0.57 seconds to first audio; only Raon SpeechChat and Deepslate Opal are listed ahead of it on that last figure, at 0.04 and 0.44 seconds.

ElevenLabs lists both models at $0.08 per minute plus LLM usage on its API pricing page. In ElevenAgents a Realtime agent bills per minute plus tokens; the API bills per minute of connected audio plus tokens.

The API follows OpenAI's Realtime spec, so an existing integration moves by changing the endpoint and the model name. It connects over WebSocket only, so a WebRTC integration has to move to WebSocket first, and it runs separately from ElevenAgents. ElevenLabs' docs say a Realtime agent hands the LLM, the TTS model, the ASR provider and a cascade timeout to ElevenLabs to manage, and swapping those means running in Custom mode.

Eleven Realtime closes a run of shipping. Instant Soundtracks, published October 15, scores an uploaded video with original music and sound effects placed at the right timestamps; the Video To Music endpoint behind it takes up to 10 videos per request, capped at 200 MB and 10 minutes. Studio 4.0 on September 21 rebuilt Studio in ElevenCreative to generate video, images, voice, music and sound effects on the timeline. Eleven v4 and v4 Turbo on September 28 added a low-latency speech model that ElevenLabs says has a median inference latency of about 100ms.

For creators the pieces form one pipeline: Eleven v4 supplies the voice, Studio 4.0 cuts the picture, Instant Soundtracks scores it, Eleven Realtime talks back live.

Pick v1 Mini or v1 Max from the model dropdown in ElevenAgents, or point an existing OpenAI Realtime client at ElevenLabs' WebSocket endpoint; current per-minute rates are on the API pricing page.