ElevenLabs' Eleven Realtime: a voice agent in a single setting, and the conditions to read before adopting it

To make a voice agent talk, you normally have to choose several components: speech recognition, turn-taking detection, the language model that formulates the reply and speech synthesis. ElevenLabs presents Eleven Realtime as a new category of "speech-to-speech" models that removes these choices: according to its changelog, you select a single model and the vendor handles the language model, transcription and speech synthesis settings behind it.

Two models are offered. Eleven Realtime v1 Mini, the default model, is described as the cheapest and aimed at natural conversations. Eleven Realtime v1 Max "reasons before it answers" and, according to the announcement, targets customer support, lead qualification and multi-step processes. The announcement states that both models support more than 90 languages and can use more than 17,000 voices from the ElevenLabs library.

A note on dates: the blog post is dated 20 October 2026, whereas the matching changelog entry in the documentation is dated 19 October.

According to the announcement, Artificial Analysis independently evaluated the two models on three public tests. On Big Bench Audio, which asks 1,000 spoken reasoning questions, Max answers 98.8% of the questions correctly and Mini 75.9%. ElevenLabs also announces an average time to first speech of 0.57 seconds for Max and 0.62 seconds for Mini. Note that the changelog and the documentation present Mini as the fastest model, whereas these measurements show Max slightly faster to start speaking.

The results do not always put the same model in front. On Full-Duplex-Bench, which tests pauses, turn-taking and interruptions, Mini scores 95.8% and Max 94.1%. On tasks that involve tools, Max succeeds at 87.1% and Mini at 60.3%, which leads ElevenLabs to recommend Max for uses that rely heavily on tools. In a phone reception setting, these tools can be a diary or a business software package.

These figures are published by the vendor, and the announcement does not say in which language the tests were run. For phone reception in France, the only test that counts remains your own: real calls in French, with callers in a hurry, accents and background noise.

Eleven Realtime is not yet open to everyone. The changelog and the documentation indicate a gradual rollout by workspace: until the setting appears in an agent's interface, the model is not available for that account.

The documentation also sets compliance conditions that clearly distinguish the two models. Both are available in the environment where data is hosted in the European Union. However, Mini is not available in "Zero Retention" mode, the option in which the vendor does not keep conversation data: only Max is, in the United States and in the European Union. A business that requires this absence of retention, for example for calls from patients or sensitive customers, will therefore have to choose Max if it wants Realtime mode, which the documentation says costs more per token than Mini, or stay in classic mode, where it chooses each component itself.

The documentation lists several restrictions. An agent that relies on its own language model, or on Amazon Bedrock, cannot switch to Realtime mode. An agent in Realtime mode can only hand the conversation over to other agents that are also in Realtime mode, Mini or Max, and can only receive it from those. This rule concerns the handover from one AI agent to another; the page devoted to Realtime models does not cover transferring a call to a person, which must therefore be checked separately.

As for service continuity, the documentation provides that if the model does not respond in time or returns an error, the agent falls back on backup language models. For developers, ElevenLabs also offers an interface compatible with OpenAI's for real-time voice; the announcement specifies that it works separately from the agents platform, and that channels such as the telephone, the knowledge base or monitoring belong to that platform.

The documentation describes billing in two parts: the usual per-minute rate of ElevenLabs' agents platform, to which are added tokens, that is, the units of text processed by the model, for input, cached input and output, at rates specific to the chosen Realtime model. Max costs more per token than Mini. The page refers to the vendor's pricing grid for the amounts.

For a small business, the real cost therefore depends on call length, the amount of text exchanged and the model chosen, rather than on a single per-minute price. Before comparing with another solution, the simplest approach is to ask for an estimate based on your actual call volume.

At Alyas, the voice agent announces at the start of the call that it is an AI, answers from the information you provide, can offer to transfer the call to the number you have configured and books appointments in Google Calendar or Cal.com once connected, with data hosted in the European Union.