ElevenLabs bets on faster and more accurate voice agents

ElevenLabs launches Eleven Realtime v1, a family of speech-to-speech models available in Mini and Max. Both are available now, compatible with OpenAI’s Realtime API and designed for voice conversations in production.

On October 20, ElevenLabs presented two real-time voice models, Eleven Realtime v1 Mini and Eleven Realtime v1 Max, designed to enable end-to-end voice agent dialogues without stacking multiple AI components. The Polish-origin startup promises low latency, better interruption handling, and support for 90 languages, with immediate access via its platform and an API compatible with OpenAI. In a market where every half-second counts, the message is clear: voice is finally fast enough to move beyond prototype.

Two models, two uses, one promise of fluidity

Eleven Realtime v1 Mini targets simple conversations where speed is paramount, while Eleven Realtime v1 Max focuses on more demanding tasks, such as customer service or agent workflows with tools. Both models were trained as true native voice agents, avoiding the classic speech-to-text then LLM then text-to-speech detour. In practice, ElevenLabs aims to replace a chain of components with a single speech-to-speech model, more direct and simpler to deploy.

ElevenLabs’ library remains available, with 17,000 voices announced. Mini can detect when the user has finished their sentence, including hesitations like “mm-hmm,” which helps maintain a natural conversation. Max emphasizes reasoning and accuracy when the request becomes less trivial.

High scores, but not on all criteria

Measures shared by ElevenLabs and the independent evaluation by Artificial Analysis place Max at the top of the ranking on several criteria. The model scores 98.8% on Big Bench Audio, versus 75.9% for Mini, and 87.1% on tasks with tool calls, versus 60.3% for Mini. However, Mini performs slightly better on conversational dynamics, with 95.8% versus 94.1% for Max. Dialogue finesse does not always obey the largest model.

Regarding latency, Max responds in 470 milliseconds and Mini in 500 milliseconds, according to published figures. For comparison, Grok Voice 3.0 is given at 810 milliseconds and GPT-Realtime-2 High at 1.14 seconds. In Artificial Analysis’s table, Max thus combines high accuracy and speed that place it in the sought-after corner for companies: fast, but not rushed to the point of saying nonsense.

An OpenAI-compatible API to ease migrations

The launch comes with a Realtime API available via WebSocket, with announced support for ElevenAgents, web, WhatsApp, and contact center platforms. WebRTC is not supported at launch. ElevenLabs emphasizes compatibility with OpenAI’s Realtime API: to migrate a voice workflow, it would be enough to change the endpoint and model name without rewriting the entire SDK.

Pricing is charged per minute, with an audio-connected cost plus tokens via the Realtime API. However, the exact token pricing was not yet communicated at the time of publication [to be verified]. The company also reserves differentiated pricing between Mini and Max, Max being more expensive than Mini according to the pricing page mentioned by the source.

Why ElevenLabs is pushing its pieces now

This release comes as ElevenLabs strengthens its presence with major European clients such as Deutsche Telekom, Revolut, and Klarna. According to provided data, large organizations now represent 55% of its revenue, up from 40% twelve months earlier. The shift is not anecdotal: it shows that AI voice is moving from a demo gadget to a production channel, where customer support, commercial qualification, or technical assistance are at stake.

In this context, ElevenLabs is not just selling a faster model. The company sells a promise of simplification, with less integration and more deployable use cases at scale. CIOs will likely appreciate the argument; developers will first look at latency. And the budget, as always, will have the final word.

Key points

  • ElevenLabs launches Eleven Realtime v1 Mini and Max.
  • Max scores 98.8% on Big Bench Audio.
  • Mini responds in 500 milliseconds according to Artificial Analysis.
  • OpenAI compatibility simplifies workflow migration.
  • Large organizations account for 55% of revenue.
  • Launch announced on October 20, 2026.

By the numbers

  • 90 languages — support announced by ElevenLabs at launch.
  • 17,000 voices — library accessible with the new models.
  • 470 ms — Max response time, according to Artificial Analysis.
  • 500 ms — Mini response time, according to Artificial Analysis.
  • 55% — share of revenue carried by large organizations.

To read