ElevenLabs has released two speech-to-speech models under the Eleven Realtime name, offering developers a choice between a lower-cost system for natural conversation and a larger model intended for more demanding tasks. Both are available through the company’s ElevenAgents platform and an OpenAI-compatible Realtime API.
The smaller model, v1 Mini, is designed for conversational applications where cost and the flow of dialogue are priorities. The larger v1 Max targets more complex interactions, including customer support. The release gives developers two options within the same model family rather than a single model for every voice application.
ElevenLabs reports that v1 Max achieves a score of 98.8 percent on the Big Bench Audio benchmark, with a response time of 0.57 seconds. The company also says that no model currently on the Artificial Analysis leaderboard is both faster and more accurate. That is a comparative claim about the combination of speed and accuracy, rather than a claim that Max leads each measure independently.
Those measurements address two important requirements for conversational AI: responding correctly and responding quickly enough to maintain a usable exchange. Benchmark performance, however, is not a complete measure of how a voice agent will behave in a deployed application. Background noise, accents, interruptions and the demands of a particular task can all affect the experience, while connections to external services can introduce additional delays.
For v1 Mini, ElevenLabs highlights the ability to distinguish between a speaker finishing a turn and making a brief acknowledgment such as “mm-hmm.” That distinction is important in spoken dialogue because an acknowledgment does not necessarily mean the listener wants to take over the conversation. A system that treats every short sound as a new turn can interrupt its own response or make an exchange feel disjointed.
Turn-taking is a broader technical challenge for voice assistants. Unlike a text chat, where sending a message supplies a clear boundary, spoken interaction requires a system to decide when a pause signals completion and when the speaker is simply hesitating. Waiting too long makes an assistant feel slow; responding too early risks cutting someone off. Mini’s stated focus places that conversational behavior alongside the usual considerations of model capability and price.
Access through an OpenAI-compatible Realtime API provides an integration route for developers using that interface format. Compatibility does not by itself establish that every existing application can switch models without changes; developers still need to check the features and behavior their implementation requires. ElevenAgents provides the other announced route to the models.
Pricing combines per-minute charges with token costs, and v1 Max is more expensive than v1 Mini. That structure makes both conversation duration and token usage relevant to estimating operating costs. ElevenLabs lists the full rates on its pricing page, leaving developers to weigh Max’s reported performance against Mini’s lower-cost positioning for their particular workloads.
Sources: The Decoder