ElevenLabs puts a managed voice stack behind Realtime Mini and Max

One setting still contains several models

The compatibility guide describes an optimized cascade: separate models handle recognition, reasoning and speech, with the components tuned to work together. This is not a single fused model. In ElevenAgents, Realtime mode lets ElevenLabs manage the model choices while developers keep control of the prompt, voice, tools, knowledge base and turn settings.

The launch offers more than 90 languages and access through ElevenAgents or a separate Realtime API. Its original announcement introduces the packaged system. The approach shares a practical question with Mistral’s new voice stack: how much setup can a provider absorb without hiding the decisions a particular service needs?

A pleasant exchange and a completed task are separate results

Artificial Analysis’s current results put Max ahead on task completion, while Mini has the higher conversational-dynamics score. Its Speech Agent Arena checks whether eligible conversations end with the correct task-completing tool calls. The timing test instead covers pauses, turn-taking, interruptions and short acknowledgements. These percentages measure different things.

The Arena’s task-success intervals are 82.6–90.6% for Max and 54.3–66.0% for Mini. Its human-preference ratings are much closer, with overlapping uncertainty intervals. A voice people enjoy is not automatically the one that carries out the request most reliably.

A useful refusal is still only a first look

Julian Goldie, whose profile promotes AI training and SEO services, reports a mock dress-order call in which the agent declined to expedite an order already in transit and offered human escalation. His post explicitly describes one call with a mock order. It is a useful example of the behaviour to inspect, with no published repeated-trial result or complete configuration for a matched comparison.

Eman Khoubian, whose profile links voice-AI business WhisperAI, asks how the headline response time holds up when a caller interrupts over a poor connection. That is a relevant test question, not a measured Realtime failure. Artificial Analysis defines its first-audio timing over spoken reasoning questions; it does not establish latency on every telephone line.

The easier setup has boundaries

Access is rolling out by workspace. The model documentation supports both Mini and Max in the EU data-residency environment, but only Max supports Zero Retention Mode there. Agents using custom language models or certain workflow overrides must remain in Custom mode; transfers between Custom and Realtime agents are restricted.

The separate API uses WebSocket, not WebRTC, and does not automatically include ElevenAgents’ platform features. The migration guide also requires browser applications to handle their audio transport. An endpoint change is not the whole migration for every existing application.

The API price list shows $0.080 per additional connected-audio minute for either model, plus tokens. Per million tokens, Mini lists $0.20 input and $6 output; Max lists $1.80 input and $9 output. Plan allowances, burst usage and the separate ElevenAgents billing route matter. Compare a configured call and its bill, rather than treating a headline minute rate as a complete operating cost.