ElevenLabs launches two voice-agent models, splitting conversation from complex tasks

ElevenLabs is promising less setup work for voice agents—not that smooth conversation alone will get the job done. On October 20, it released Eleven Realtime v1 Mini and v1 Max through ElevenAgents and a developer API. Its published tests separate the models’ strengths: Mini leads slightly on conversational dynamics, while Max performs better on reasoning and tool-use tasks.

One model choice, several jobs

A voice agent has to recognize speech, decide when someone has finished talking, work out an answer and speak it aloud. Eleven Realtime combines speech recognition, turn-taking, a large language model and text-to-speech, with those components tuned together as one system.

The company’s pitch is to reduce the work of choosing and tuning those parts separately. That includes small but consequential decisions: how long to wait after a pause, how to recover from an interruption and how a voice works with the agent’s instructions. ElevenAgents still lets builders adjust settings, but the new models offer a preconfigured starting point.

Mini is the lower-cost option for assistants, tutors and interactive characters. Max is aimed at customer support, lead qualification and workflows requiring several steps. It adds reasoning capability for following a process and calling connected software tools. Both models support more than 90 languages and voices from ElevenLabs’ library of more than 17,000.

The tests measure different kinds of success

ElevenLabs says Artificial Analysis independently evaluated both models across three public benchmarks: Big Bench Audio, Full-Duplex-Bench and Speech Agent Arena. The results published in its launch announcement distinguish answering spoken questions, handling conversational timing and completing tasks. Those are separate measures, rather than one overall score for a voice agent.

Big Bench Audio asks 1,000 reasoning questions by voice and scores the spoken answers. Max answered 98.8% correctly, compared with Mini’s 75.9%. Max also began speaking sooner in that test: an average of 0.57 seconds, versus 0.62 seconds for Mini. These figures measure time to first audio, not the time needed to finish an answer or complete a customer request.

Mini’s advantage appeared in Full-Duplex-Bench, which tests pauses, turn-taking, interruptions and brief listener acknowledgments. It scored 95.8%, against Max’s 94.1%. In Speech Agent Arena, where people compare models in blind live conversations, Max scored 998 and Mini 988. The conversational-dynamics result and the human-preference score therefore do not produce the same ordering.

The larger separation came on tool-use tasks: Max completed 87.1%, while Mini completed 60.3%, according to ElevenLabs. That is the basis for its recommendation to use Max when an agent relies heavily on connected tools. Even Max’s stronger result falls short of completing every tested task; a high speech-reasoning score is not a task-completion guarantee.

Two access routes, with different boundaries

In ElevenAgents, builders can open a new or existing agent and select Mini or Max from a dropdown. Existing prompts, tools, knowledge bases, channels and settings remain available. Existing agents do not change unless their builders update them. The platform also supplies deployment channels including phone, web and WhatsApp, alongside guardrails, testing and monitoring.

For developers using OpenAI’s Realtime API, ElevenLabs offers a compatible interface that works with OpenAI’s software development kit. Its Realtime product page describes keeping existing event handlers, logging and tooling. But compatibility comes with two important boundaries:

• The API connects over WebSocket only. An existing WebRTC integration must switch connection methods first; changing the endpoint and model name alone will not cover that migration.
• The API runs separately from ElevenAgents. Builders needing the platform’s phone or WhatsApp channels, knowledge base, guardrails, testing or monitoring must build on ElevenAgents instead.

Billing also has two components. ElevenAgents charges its standard per-minute rate plus language-model token costs—the charges for units of model input and output. Direct API use is billed per minute of connected audio plus tokens, with the same token prices across plans. The agent-model selector shows an estimated per-minute cost, while the product page directs customers to current pricing for rates.