The Pulse
ElevenLabs Launches Realtime Speech-to-Speech Models for Voice Agents
ElevenLabs launched Eleven Realtime v1 Mini and v1 Max for building voice agents, with support for more than 90 languages and 17,000-plus voices. The company says Max scored 98.8% on a speech-reasoning benchmark, while both models are avail
AI.info Team · Written with AI · fact-checked ·
ElevenLabs launched two speech-to-speech models on October 20, 2026, aimed at developers building voice agents. The company says Eleven Realtime combines speech recognition, turn-taking, a language model and voice output in a single agent choice, instead of asking builders to configure each component separately. Min Kim, the author of ElevenLabs’ announcement, writes: “Each model is a complete voice agent, optimized for separate use cases.”
Mini and Max target different conversations
Eleven Realtime v1 Mini is designed for lower-cost, natural exchanges such as AI assistants, tutors and interactive characters. The v1 Max model is aimed at customer support, lead qualification and other workflows where an agent must follow steps, use tools or reason through a more complex request. Both models support more than 90 languages and can use voices from ElevenLabs’ library of more than 17,000 voices, according to the company.
The model choice changes how the agent’s components are managed, but does not remove every configuration option. ElevenLabs says prompts, voices, tools, knowledge bases and conversation-flow settings remain available in Realtime mode. Its documentation also says Realtime models are for voice agents created in ElevenAgents; agents built through other ElevenLabs products, including Speech Engine, cannot use them.
Benchmark results favor Max on reasoning and tools
ElevenLabs says Artificial Analysis independently evaluated the models on three public benchmarks. In Big Bench Audio, which asks 1,000 spoken reasoning questions and scores the answers, Max answered 98.8% correctly and Mini answered 75.9%. The company reports average time to first audio of 0.57 seconds for Max and 0.62 seconds for Mini.
The results are not uniformly higher for Max. On Full-Duplex-Bench, which tests pauses, turn-taking, interruptions and backchannels, Mini scored 95.8% and Max 94.1%. ElevenLabs reports Speech Agent Arena scores of 998 for Max and 988 for Mini, and tool-use task completion rates of 87.1% and 60.3%, respectively. Those figures come from the company’s launch post; they describe different evaluations and should not be read as a single overall score.
OpenAI-compatible API comes with a transport limit
Developers can select a Realtime model inside ElevenAgents or connect through ElevenLabs’ Realtime API. The company says the API follows OpenAI’s Realtime specifications, allowing developers with an existing compatible integration to change the endpoint and model name rather than rebuild the client. The connection uses WebSocket, however, so an integration built for WebRTC must move to WebSocket to use this API.
The Realtime API also runs separately from ElevenAgents. Developers who need ElevenAgents features such as phone and WhatsApp channels, workflows, a knowledge base, guardrails or monitoring must use that platform rather than the standalone API, according to the announcement.
Pricing and rollout are not identical for both models
ElevenLabs’ documentation says Realtime agents carry the standard ElevenAgents per-minute charge plus input, cached-input and output token charges at the model’s rates; Max costs more per token than Mini. The launch post does not provide specific prices. The documentation also describes availability as a gradual rollout, so the Realtime option may not yet appear in every workspace.
That leaves developers weighing two practical constraints alongside the benchmark results: whether Realtime has reached their workspace and whether their existing audio client uses WebSocket. Mini is the lower-cost, lower-latency choice in ElevenLabs’ product guidance; Max is the more expensive option for complex instructions and multi-step tool use.