ElevenLabs Realtime: Realtime voice agent models with 570ms to first audio

Mini reaches first audio in 620ms, and Max is the faster model at 570ms, according to figures ElevenLabs credits the evaluations to Artificial Analysis. On Big Bench Audio, a set of 1,000 reasoning questions asked by voice, Max scores 98.8% against 75.9% for Mini. Max also completes tool-use tasks 87.1% of the time, versus 60.3% for Mini. Mini holds a small lead on conversational flow, at 95.8% for Mini and 94.1% for Max on Full-Duplex-Bench. Listed pricing for both tiers is $0.08 per minute plus LLM usage. The models went live on October 20, 2026, across ElevenAgents and the Realtime API. They support 90+ languages. ElevenLabs says no model on the Artificial Analysis leaderboard is both faster and more accurate. Against OpenAI, Mini's conversational dynamics score is on par with OpenAI's GPT-Realtime 2.1. The latency and accuracy numbers are vendor-reported, so independent verification is still pending. For founders already on OpenAI's Realtime stack, migration is mostly configuration: the endpoint and model name change while the rest of the code stays the same. The tradeoff is control. Developers cannot select or separately configure the underlying models. Compliance terms also differ by tier, since Mini lacks Zero Retention Mode and HIPAA support, features Max offers in the US and EU but not in India or Singapore. Teams should run their own call recordings through both models before moving production traffic.