ElevenLabs Launches Eleven Realtime, Voice Models That Reply in About Half a Second

ElevenLabs has released two new voice models called Eleven Realtime, built for AI conversations that feel like a real phone call. According to The Decoder on October 20, there are two versions. v1 Max is aimed at harder jobs such as customer support and scores 98.8% on the Big Bench Audio test with a response time of 0.57 seconds. v1 Mini is designed for natural everyday conversation at lower cost. ElevenLabs says it can tell when a speaker has really finished, as opposed to just saying “mm-hmm” to show they are listening.

These are “speech-to-speech” models. Older voice assistants chain three steps: turn your speech into text, let a chatbot think about the text, then turn the answer back into speech. Each step adds delay. A speech-to-speech model works with sound more directly, which is one way to cut the awkward pause and to catch tone and timing better. The Decoder’s chart places Max near the top left, meaning high accuracy and low delay, and ElevenLabs claims no model on the Artificial Analysis leaderboard is both faster and more accurate. Both models are available through the ElevenAgents platform and an OpenAI-compatible Realtime API. Another report adds support for more than 90 languages.

Treat the claims with some care. The speed and accuracy figures come from ElevenLabs, and the Decoder article gives no results for Mini, no list of languages, and no detailed limitations. Pricing is per minute plus token costs, with Max costing more than Mini, but specific rates are only on the company’s pricing page.

What’s behind it: a “benchmark” is a standard test used to compare models, and Big Bench Audio checks how well a model handles spoken questions. A high score is a good sign but is not the same as handling a noisy line, strong accents or an angry customer. Response time matters because humans are sensitive to pauses in conversation. Around half a second feels natural; two seconds feels broken. That is why so many companies, including Mistral this week with Voxtral 3, are racing to make voice agents quicker and cheaper.

What this means for you: if you use voice assistants occasionally, expect them to sound smoother over the coming months. If you build customer support or booking tools, this is another option to test next to Mistral’s and others. Try it with your own recordings and your real customers’ accents before you decide, and read the pricing page, because per-minute costs add up quickly at scale.

Sources

• The Decoder: ElevenLabs releases two new speech-to-speech models for real-time AI conversations
• The Neuron: Everything that happened in AI today, October 21