ElevenLabs folds the whole voice-agent stack into one model, and its Max tier answers 98.8% of spoken reasoning questions
Synopsis
ElevenLabs released Eleven Realtime, a family of speech-to-speech models that each act as a complete voice agent, replacing separate choices for speech recognition, voice activity detection, turn-taking, the underlying LLM and text to speech. Both support more than 90 languages and the library's 17,000+ voices. In independent evaluation, v1 Max answered 98.8% of 1,000 spoken reasoning questions correctly and began speaking in 0.57 seconds on average; v1 Mini scored 75.9% and 0.62 seconds.
Interpretation
Building a natural-sounding voice agent shifts from assembling components to picking one complete model: each Eleven Realtime model is itself a voice agent, optimized for separate use cases, and supports more than 90 languages plus 17,000+ voices from the ElevenLabs library. Previously, developers had to decide on speech recognition, voice activity detection, turn-taking, the underlying LLM and text to speech separately, then tune them to work together, which was a project in itself. Artificial Analysis evaluated both v1 Mini and v1 Max independently across three public benchmarks: Big Bench Audio, Full-Duplex-Bench and the Speech Agent Arena.
The two tiers split by task complexity: v1 Mini targets low-cost natural conversation for AI assistants, teachers and interactive characters, while v1 Max targets complex tasks that follow a process and reason through steps, such as customer support and lead qualification. Accuracy and speed are usually seen as a trade-off, with the highest-scoring models on Big Bench Audio taking longest to answer and the fastest giving up accuracy. On the 1,000 spoken reasoning questions of Big Bench Audio, v1 Max answered 98.8% correctly and started speaking in 0.57 seconds on average, while v1 Mini answered 75.9% and started in 0.62 seconds; on Full-Duplex-Bench, v1 Mini scored 95.8% and v1 Max 94.1%.
Conversational dynamics and human preference are measured separately: in the Speech Agent Arena, where people compare two models in blind live conversations, v1 Max scored 998 and v1 Mini 988; on tool-use tasks, v1 Max completed 87.1% and v1 Mini 60.3%. Details such as how long to wait after a pause before answering and how to recover when the user interrupts previously depended on developer-adjusted settings. Full-Duplex-Bench tests pauses, turn-taking, interruptions and backchannels, while the Speech Agent Arena scores human preference and task success in blind live conversations.
Perspective
The release targets developers building voice agents, with one tier for low-cost natural conversation and one for complex multi-step tasks. The Realtime API connects over WebSocket only, so a WebRTC integration must move to WebSocket first; that API runs separately from ElevenAgents, so teams needing channels or a knowledge base should build on ElevenAgents from the start.
The accuracy-versus-speed trade-off is described by the vendor, and the claim that v1 Max leads on both metrics comes from the evaluation it cites. On tool use, v1 Mini completes 60.3% versus 87.1% for v1 Max, which is why Max is recommended for heavy tool calling. The two tiers score close on Full-Duplex-Bench, with v1 Mini slightly higher.