ElevenLabs has launched Eleven Realtime, a family of speech-to-speech models built for voice agents. The release includes two versions: Eleven Realtime v1 Mini and Eleven Realtime v1 Max, each designed for different types of applications.
Both models support more than 90 languages and can use any of the more than 17,000 voices in the ElevenLabs library. That gives developers a broad set of options for creating voice agents, whether the goal is an AI assistant, a teacher, an interactive character, or a system that handles customer support.
“You shouldn’t have to become a configuration expert to build a voice agent that sounds natural,” the company said. The two-model lineup separates everyday voice interactions from applications that depend on tool use and multi-step workflows.
Mini and Max target different voice-agent jobs
Eleven Realtime v1 Mini targets applications such as AI assistants, teachers, and interactive characters. It is also built for live conversations, with an average time to first audio of 0.62 seconds in independent evaluations by Artificial Analysis.
Eleven Realtime v1 Max is aimed at customer support, lead qualification, and workflows involving multiple steps. ElevenLabs recommends Max for applications that rely heavily on tool calling and multi-step workflows, where the voice agent must complete actions as part of a conversation.
The performance results show a clear difference between the models in several tests. On Big Bench Audio, Max answered 98.8% of questions correctly, compared with 75.9% for Mini. Max also recorded an average time to first audio of 0.57 seconds, compared with 0.62 seconds for Mini.
Mini led on one test focused on handling two-way conversation. It scored 95.8% on Full-Duplex-Bench, while Max scored 94.1%. In the Speech Agent Arena, Max scored 998 and Mini scored 988.
Max leads on reasoning and tool use
The biggest gap appears in tool-use tasks, which measure how well a model completes actions through connected tools. Max completed 87.1% of these tasks, compared with 60.3% for Mini.
That result helps explain the different roles ElevenLabs gives each model. Mini fits voice experiences where natural interaction is the main goal, while Max fits systems that must answer questions, call tools, and move through several steps before completing a workflow.
The Big Bench Audio results also put the models on different levels for speech reasoning. Max reached 98.8% accuracy, while Mini reached 75.9%. At the same time, Mini’s higher Full-Duplex-Bench score shows that the smaller model has a strong result in a test focused on simultaneous conversational exchange.
These results do not make one model the right choice for every application. Instead, they give developers a way to match the model to the work a voice agent needs to perform. A teacher or interactive character may use Mini, while customer support or lead qualification may call for Max.
Existing OpenAI-compatible integrations can move over
ElevenLabs says its models are compatible with OpenAI’s Realtime API and follow OpenAI’s Realtime specifications. Developers can switch to ElevenLabs by changing the API endpoint and model name in their existing integrations.
The API connects through WebSockets and supports OpenAI’s SDK. Developers using WebRTC integrations will need to move to WebSockets to use the service, so the connection method is an important part of the transition.
For teams that already work with OpenAI’s Realtime API, the endpoint and model-name changes offer a direct path to testing Eleven Realtime. The WebSockets requirement still means WebRTC integrations need a different connection setup before they can use the service.
Developers can also access the models through ElevenAgents, the company’s agent platform. ElevenAgents provides deployment across multiple channels, giving teams another way to put Eleven Realtime models into voice-agent applications.
Eleven Realtime therefore arrives as more than a single voice model. It offers two versions, support for more than 90 languages, access to more than 17,000 voices, and compatibility with OpenAI’s Realtime API and SDK. Mini focuses on live conversational uses, while Max is built for reasoning, tool calling, and workflows with several steps.
