ElevenLabs releases Eleven Realtime for faster speech-to-speech AI

ElevenLabs has released two speech-to-speech models under the Eleven Realtime name, according to The Decoder. The launch matters for teams building voice agents and conversational apps because the company is putting speed, accuracy and turn-taking at the center of the product.

The release includes a smaller v1 Mini model and a more capable v1 Max model. Both are available through ElevenLabs’ ElevenAgents platform and an OpenAI-compatible Realtime API, giving developers and product teams more than one route to test the models in voice workflows.

ElevenLabs is already known as an AI voice platform, and this release pushes further into live, two-way speech rather than one-way voice generation. For creators, marketers and small businesses, the practical question is less about synthetic narration and more about whether voice interfaces can respond quickly enough to feel conversational.

What Eleven Realtime adds

Eleven Realtime is not a single model. ElevenLabs is shipping two variants aimed at different trade-offs.

The v1 Mini model is designed for more natural conversations at a lower cost, according to the company. The Decoder reports that it can distinguish between a person finishing a thought and a brief backchannel sound such as “mm-hmm,” which is important in spoken interfaces because interrupting too early can make an AI agent feel clumsy.

The v1 Max model is aimed at more complex uses. ElevenLabs points to customer support as one target use case, which suggests the company sees this model as a fit for higher-stakes conversations where comprehension and timing both matter.

According to the report, v1 Max scored 98.8 percent on the Big Bench Audio benchmark and has a 0.57-second response time. ElevenLabs says that, at the moment, no model on the Artificial Analysis leaderboard is both faster and more accurate.

Why latency and turn-taking matter

For real-time voice apps, a model’s voice quality is only one part of the experience. The timing of a reply often determines whether a system feels like a conversation or like a delayed voice menu.

That is why the response-time claim around v1 Max is notable. A faster speech-to-speech model can reduce pauses between a user speaking and the AI answering. For customer support, intake calls, coaching flows or interactive demos, those pauses can affect whether users keep talking naturally or start adapting their speech to the machine.

The Mini model’s focus on knowing when a speaker has finished is also relevant. Spoken conversations include pauses, fillers and acknowledgements that do not always mean the user is done. If an AI agent treats every hesitation as an endpoint, it can talk over people; if it waits too long, the exchange feels slow. ElevenLabs is explicitly positioning Mini around that conversational edge case.

For teams comparing the broader voice-agent market, this launch sits alongside a wave of new real-time audio systems. We recently covered Mistral AI’s Voxtral 3 for real-time voice agents, and buyers can also use our Best AI Voice Generators in 2026 guide as a wider starting point.

How teams can access the models

ElevenLabs says both Eleven Realtime models are available through ElevenAgents and an OpenAI-compatible Realtime API. That gives no-code or agent-platform users one path, while developers can work through an API route that follows a familiar compatibility pattern.

Pricing is based on minutes plus token costs, according to the report. ElevenLabs says v1 Max costs more than v1 Mini, and directs users to its pricing page for full details.

For small teams, that pricing structure makes model selection an operational decision. Mini may be the first model to test where cost and conversational feel are the main concerns. Max is the option ElevenLabs is positioning for more complex tasks, including customer support.

Because the source does not provide deployment limits, supported languages or detailed pricing, teams should verify those details directly before planning production rollouts.

What to watch next

The immediate benchmark claim is clear: ElevenLabs says v1 Max combines a 98.8 percent Big Bench Audio score with a 0.57-second response time, and that no current Artificial Analysis leaderboard model is both faster and more accurate. The more important test will be how those numbers translate into real user conversations.

For creators and marketers, the opportunity is in interactive voice experiences: guided product explainers, spoken lead qualification, training simulations or voice-based support flows. For freelancers and agencies, the useful experiment is to compare Mini and Max on the same script and conversation pattern, then decide whether the additional capability of Max is worth the higher cost.

The release also raises the bar for anyone building voice agents. Real-time speech systems now have to compete not only on voice realism, but on interruption handling, response timing and the ability to keep a conversation moving without awkward gaps.