ElevenLabs released the Eleven Realtime speech-to-speech model family on October 20, 2026, aiming to retire the pipeline voice agents have relied on for years: speech recognition, then a language model, then speech synthesis, with latency and errors accumulating at every seam. Eleven Realtime replaces that chain with a single native voice model that listens and answers directly. There are two variants: Eleven Realtime v1 Mini, tuned for simple conversations at the lowest latency, and Eleven Realtime v1 Max, tuned for reasoning and accuracy on harder tasks. Both support up to 90 languages, can draw on the 17,000 voices in the ElevenLabs library, and are trained for human conversational dynamics such as pauses, interruptions and turn-taking.
The split between the two models matters more than the names
Mini targets high-volume, simple dialogue: questions, bookings and status checks where little deliberation is needed; its conversational naturalness even scores slightly above Max. Max is built for work that requires thinking: customer support, agent workflows, multi-step reasoning and calling tools mid-conversation. Selling the two separately is itself a signal: the voice-agent market has outgrown one-model-fits-all, and latency-sensitive work is a different business from reasoning-sensitive work.
ElevenLabs has been pushing into real-world settings, as when an ElevenLabs voice clone addressed the House of Lords, which showcased synthesis quality. Realtime addresses the other half of the problem: however lifelike the voice, a slow or wrong answer still makes the person on the phone hang up.
Reading the benchmarks: fast and accurate, for once, together
On the Big Bench Audio voice-reasoning benchmark by Artificial Analysis, as cited by ElevenLabs, Max scores 98.8%, just ahead of Grok Voice 3.0 at 98.5% and Gemini 3.8 Live Extended Thinking at 97.7%; Mini scores 75.9%. On latency, Max starts responding in about 470 milliseconds and Mini in about 500, versus 810 milliseconds for Grok Voice 3.0 and 1.14 seconds for GPT-Realtime-2 (High). Tool-calling success splits the family: 87.1% for Max against 60.3% for Mini. Conversational dynamics score 95.8% for Mini and 94.1% for Max, on par with OpenAI's GPT-Realtime 2.1, and in the blind head-to-head Speech Agent Arena, Max scores 998 and Mini 988.
These are vendor-published figures measured on the model itself, excluding telephony networks, queuing under load and round trips to business systems. Live deployments routinely run slower end to end, so the table is a shortlist, not a specification.
Choosing: the 27-point tool-calling gap decides
For high-frequency, short, deterministic voice entry points, Mini's trade-off is right: cost stays on simple dialogue, naturalness does not suffer, and a half-second response keeps calls from going cold. If the agent must check orders, change bookings or walk through multi-step processes while talking, the tool-calling gap of 87.1% versus 60.3% is what counts; one wrong action costs far more than 30 extra milliseconds.
For teams building voice support and phone agents, the deeper change is architectural: a native speech-to-speech model no longer flattens tone, emotion and interruptions through a text relay, and debugging converges from three pipelines to one model. The practical test is simple: blind-test the 50 most common real questions from your own traffic, record high-percentile first-response latency and tool-call success, and discount the leaderboard accordingly. Benchmarks should draw up the shortlist; contracts should be signed on your own measurements.
