Inside this story Context & sources

This article links to 2 sources. Source links are provided below so you can inspect the evidence yourself.

  • ElevenLabs launched Eleven Realtime Mini and Max on 20 October, targeting different kinds of spoken interaction.
  • Artificial Analysis measures response speed, task success and conversational behaviour separately.
  • A useful local pilot would check a whole exchange in the intended languages and speaking patterns.

A conversation has several moving parts

ElevenLabs launched Eleven Realtime on 20 October, introducing v1 Mini and v1 Max as speech-to-speech models. The company describes each as a complete voice agent, bringing more of the listening and speaking process into one offering. Its pitch is less assembly work for people building a spoken assistant. Whether the resulting conversation feels useful still needs a closer listen. [1]

The distinction between the two versions concerns the job. ElevenLabs positions Mini for natural conversation, including assistants and interactive characters, and Max for more involved tasks and multi-step workflows. It says both support more than 90 languages. That broad coverage is a starting point for evaluation; it does not establish how either handles a particular South African language, accent or mixed-language exchange. [1]

A quick reply is one measure

Artificial Analysis provides a useful second view through its own speech-to-speech leaderboard. Its table lists average time to first audio of 0.57 seconds for Max and 0.62 seconds for Mini. The measure concerns the beginning of the spoken response on its Big Bench Audio test. It is not a timing for an entire conversation or a promise about a caller’s connection. [2]

The evaluator also lists task success rates of 87.1% for Max and 60.3% for Mini in its Speech Agent Arena. That measure asks whether eligible conversations ended with the correct task-completing tool calls. It measures something different from how quickly speech starts. These are published evaluation results, not tests performed by The Open Tab or a forecast for a particular business. [2]

Leave room for the pause

Artificial Analysis assesses conversational behaviour separately, including whether a model waits through a pause and responds appropriately to an interruption. Our reading is that this deserves attention alongside speed. A person may be looking for a reference number, correcting a detail or simply taking a breath. A reply that arrives quickly can still arrive at the wrong moment. [2]

For a South African team exploring spoken AI, we would start with a short, repeatable practice conversation using invented details. Include a pause, a correction and one ordinary request that needs a clear result. Try the languages and speaking patterns the intended users would actually choose. Record whether the result is right and whether the exchange leaves enough room for the speaker.

Choose the route as well as the model

ElevenLabs offers the new models through ElevenAgents and a separate Realtime API. Its launch post distinguishes the platform’s channels and knowledge-base features from the separate API route. That is a useful question to settle before a pilot: which surrounding capabilities does the proposed experience need? [1]

The practical opportunity is a simpler starting point for a voice prototype. We would judge the outcome through a complete exchange, with a clear task and space for the person to speak. We have not run this proposed exercise, tested the models or verified local calling costs. The launch and the evaluator’s results give a reason to investigate, rather than a finished verdict.

Sources & further reading

1. ElevenLabs: Introducing Eleven Realtime — 20 October 2026

Original dated launch by Min Kim. Mini/Max positioning, language coverage and product routes are vendor statements. No customer outcome, compatibility test or local performance inferred.

2. Artificial Analysis: Speech to Speech leaderboard and methodology

Original evaluator results and methodology, read on 21 October. The timings measure the start of audio; task success concerns eligible Arena conversations. No publication or evaluation date is established for this live page, and these results do not establish South African performance.

AI-assisted research and writing, with a separate source-checking pass by the same Codex assistant. Product descriptions remain attributed to ElevenLabs; Artificial Analysis supplies its own published evaluation and methodology, not a local deployment test. No South African language, accent, calling-cost or hands-on results have been verified. The suggested practice exercise has not been performed by us.

THE OPEN TABStay curious. Keep reading.