• Verdict: Wait. The leading number is real enough to test, but it comes from one benchmark and one latency metric.
• The Artificial Analysis speech-to-speech page lists Eleven Realtime v1 Max at 99% on Big Bench Audio with 0.57 s to first audio, consistent with the 98.8% ElevenLabs reports.
• The cheaper v1 Mini is listed at 76% on the same benchmark and 0.62 s, so the tier choice changes accuracy far more than speed.
• Both models are reported to run on the ElevenAgents platform and an OpenAI-compatible Realtime API; pricing is reported as per-minute plus token costs, and we have no per-call figure.
• A benchmark score says little about barge-in, noisy lines, accents or tool-call failures, which are what usually break voice agents in production.

Verdict: Wait. ElevenLabs has released two speech-to-speech models, Eleven Realtime v1 Max and v1 Mini, and reports that v1 Max scores 98.8 percent on Big Bench Audio with a 0.57-second response time. Unlike many vendor claims, these figures also appear on a third-party leaderboard. That earns the model a trial, not a migration, because one benchmark and one latency metric do not describe a production voice agent.

What the independent listing shows

The Artificial Analysis speech-to-speech page lists Eleven Realtime v1 Max at 99 percent on Big Bench Audio with 0.57 seconds to first audio. That is consistent with the 98.8 percent and 0.57 seconds that The Decoder relays from ElevenLabs, the difference being rounding. We read the listing but did not rerun the measurement, so we cannot vouch for how the latency was timed or on what network.

The tier gap matters more than the headline

The same listing shows v1 Mini at 76 percent on Big Bench Audio and 0.62 seconds to first audio. The choice between tiers therefore trades roughly 23 points of benchmark accuracy for a small latency difference. The Decoder describes Mini as built for natural, lower-cost conversation, including telling a finished turn from a backchannel such as "mm-hmm," and Max as aimed at harder work such as customer support.

Why a benchmark score is not a production verdict

A score near 99 percent on one benchmark says the model handles that test set well. It does not measure what typically fails in deployed voice agents: interruptions, noisy lines, accents, long calls, or tool calls that error mid-turn. A demo hides those cases. Treat the leaderboard entry as a reason to replay your own recordings through the model, not as a substitute for doing so.

Cost is still the missing half

ElevenLabs is reported to charge per minute plus token costs, with Max priced above Mini, and both models are reported to be available through ElevenAgents and an OpenAI-compatible Realtime API. The Decoder does not list rates, and we have no per-call figure. Without a cost per resolved conversation, the accuracy and latency numbers cannot be weighed against the cascaded speech-to-text, language model and text-to-speech pipeline many teams run today.

The bottom line

Bottom line: Wait. The deciding number is 99 percent on Big Bench Audio at 0.57 seconds, now visible on Artificial Analysis as well as in the vendor's claims. Trial v1 Max, not Mini, if accuracy matters: replay real calls, measure time to first audio at your own percentiles, count barge-in and tool-call failures, and price the result per resolved call. If it holds on your audio, the verdict moves to Use it.