ElevenLabs released Eleven Realtime, a pair of speech-to-speech models that each work as a complete voice agent. ElevenLabs says each one removes the work of choosing and tuning separate parts for speech recognition, voice activity detection, turn-taking, the language model and text to speech.
Eleven Realtime v1 Mini is the low-cost model for assistants, teachers and interactive characters, and the company expects it to be enough for most uses. v1 Max is aimed at customer support, lead qualification and multi-step workflows, where the agent has to follow a process and reason through each step. Both support more than 90 languages and can speak in any of the 17,000+ voices in the ElevenLabs library.
Artificial Analysis tested both Eleven Realtime models on three public benchmarks, ElevenLabs reports. On Big Bench Audio, which asks 1,000 reasoning questions by voice and scores the spoken answer, Max answered 98.8% correctly, while Mini answered 75.9% correctly. Max starts speaking after 0.57 seconds on average and Mini after 0.62. ElevenLabs says the top scorers on that test usually take the longest to answer and that Max does well on both counts.
On Full-Duplex-Bench, which tests pauses, turn-taking, interruptions and backchannels, Mini scores 95.8% and Max 94.1%. The third benchmark, the Speech Agent Arena, has human judges rate each model without knowing which one they hear, and there Max reached 998 against 988 for Mini. The models part ways on tool use: Max completed 87.1% of tasks and Mini 60.3%, and ElevenLabs recommends Max for agents that call tools heavily.
Beyond the minutes each plan includes, both models bill $0.08 per additional minute of connected audio plus tokens, according to ElevenLabs' API pricing. Max input tokens cost $1.80 per million, nine times Mini's $0.20. Max output tokens cost $9.00 per million against Mini's $6.00. Each self-serve plan also caps concurrent calls, at 4 to 40 depending on the plan, and connected audio above that cap bills at a $0.16 burst rate.
| Charge | v1 Mini | v1 Max |
|---|---|---|
| Connected audio, per additional minute | $0.08 | $0.08 |
| Input tokens, per 1M | $0.20 | $1.80 |
| Output tokens, per 1M | $6.00 | $9.00 |
| Burst rate above concurrency limit, per minute | $0.16 | $0.16 |
Developers can pick either model from the dropdown when they create or edit an agent in ElevenAgents, which adds phone, web, WhatsApp and contact-center channels, workflows, a knowledge base, guardrails, testing and monitoring. Eleven Realtime also runs through a standalone Realtime API that follows OpenAI's Realtime specification, and ElevenLabs says a team already on OpenAI's Realtime API can try it by changing the endpoint and the model name, with OpenAI's SDK working as is. ElevenLabs publishes a migration guide covering the differences from OpenAI's API.
The Realtime API runs separately from ElevenAgents, and ElevenLabs recommends that teams needing channels or a knowledge base build on ElevenAgents from the start. The API also connects over WebSocket only, and a WebRTC integration has to move to WebSocket first.
