News

Eleven Realtime: ElevenLabs' Speech-to-Speech Models Reach 90 Languages, With Two Sets of Latency Numbers

22 October 2026 · RenderBob team

ElevenLabs announced Eleven Realtime on 20 October, two voice-agent models called Mini and Max. The latency figures differ between the two reports we read.

ElevenLabs announced Eleven Realtime on 20 October, two voice-agent models called Mini and Max. The latency figures differ between the two reports we read, so a comparison with Grok Voice 3.0 needs care.

What was announced

Journal du Net reports that Eleven Realtime v1 Mini and Max listen, reason and reply in one managed model. Mini is meant for simple conversations where latency matters most, and Max for multi-step reasoning such as customer service and agent workflows. ElevenLabs says they support up to 90 languages and give access to the 17,000 voices in its library. They are available through ElevenAgents, which covers phone, web, WhatsApp and contact-centre platforms, and through a Realtime API over WebSocket. WebRTC is not supported at launch.

The numbers, and where they disagree

• Latency. Journal du Net gives 470 ms for Max and 500 ms for Mini, against 810 ms for Grok Voice 3.0 and 1.14 seconds for GPT-Realtime-2 (High). The New Stack, citing Artificial Analysis, gives time to first audio of about 570 ms for Max and 620 ms for Mini. We cannot tell whether these are different measurements or different snapshots.
• Benchmarks. Both outlets list the same Artificial Analysis results: Big Bench Audio at 98.8% for Max and 75.9% for Mini, tool-call success of 87.1% for Max and 60.3% for Mini, and Full-Duplex-Bench at 94.1% for Max and 95.8% for Mini.
• Speech Agent Arena. Journal du Net lists 998 for Max and 988 for Mini. We reported 1070 Elo for Grok Voice 3.0 on 16 October. We have not confirmed that the two figures are on the same scale or from the same snapshot, so do not rank them against each other.

Pricing and compatibility

In ElevenAgents the models use the standard per-minute rate. Through the Realtime API, usage is billed per minute of connected audio plus tokens, and the token price had not been announced in the coverage we read. ElevenLabs says the API is fully compatible with OpenAI's Realtime API, so moving over means changing the endpoint and model name. The New Stack adds that WebSocket users must capture and play 24 kHz PCM16 audio themselves.

What you give up

The New Stack's argument is that ElevenLabs is taking over the voice-agent harness, meaning turn detection, interruption handling and orchestration. You cannot choose the underlying models in this mode, and the two Realtime models cannot transfer calls to agents in Custom mode, where you choose your own. Mini lacks Zero Retention Mode and HIPAA configurations. Max supports both in the US and EU but is not available in India or Singapore data residency environments. ElevenLabs did not respond to that outlet's request for comment.

Why a filmmaker might care

Real-time voice models matter for interactive characters, installations and live demos, where a figure must answer a person in the moment. For that use, check the data terms before you connect: where audio goes, how long it is kept, and whether a performer's voice is involved. Our voice-release comparison from 21 October set out the same questions for the other launches.

What to watch

The token price, independent listening and latency tests, and whether ElevenLabs publishes its voice-consent policy for these models.

RenderBob 3 and this story

Eleven Realtime is for a voice that answers in the moment. RenderBob 3 is for the picture. An imported voice sample stays with the project beside the prompts for the shots. Importing a sample does not generate speech.