ElevenLabs releases Eleven Realtime v1 Max, claims 98.8% on Big Bench Audio at 0.57s response time

ElevenLabs released two speech-to-speech models, Eleven Realtime v1 Max and Eleven Realtime v1 Mini, on October 20, 2026. According to the company, v1 Max scores 98.8% on the Big Bench Audio benchmark with a 0.57-second response time.

Key specifications

Detail	Eleven Realtime v1 Max	Eleven Realtime v1 Mini
Positioning	Complex tasks such as customer support	Natural conversation at lower cost
Big Bench Audio	98.8% (per ElevenLabs)	Not disclosed
Response time	0.57 seconds	Not disclosed
Pricing	Per minute plus token costs; higher than Mini	Per minute plus token costs; lower than Max
Context window	Not disclosed	Not disclosed

Exact per-minute and per-token prices were not included in the announcement. ElevenLabs points to its pricing page for details. Parameter counts and training cutoff dates have also not been disclosed.

Performance claims

ElevenLabs says no model on the Artificial Analysis Speech to Speech leaderboard is currently both faster and more accurate than v1 Max. In the company's chart, v1 Max sits near the top-left corner of the leaderboard, which combines high accuracy with low latency. This is a vendor claim based on a chart ElevenLabs supplied, and we have not independently verified it. Big Bench Audio scores for v1 Mini were not published.

Model behavior

v1 Mini is designed for natural conversational flow. ElevenLabs says it can distinguish a speaker who has finished talking from one who is only making a backchannel sound such as "mm-hmm." Turn-taking of this kind is a common failure point in voice agents, which tend to interrupt users or respond to filler sounds. v1 Max targets more demanding workloads, with customer support named as the primary example.

Availability

Both models are available through:

• The ElevenAgents platform
• An OpenAI-compatible Realtime API

The OpenAI-compatible interface means developers with existing integrations built against that API format may be able to switch models with limited code changes. ElevenLabs has not detailed compatibility limits in the announcement.

What this means

Speech-to-speech models process audio directly rather than chaining transcription, a language model and text-to-speech. That reduces latency and preserves vocal cues. ElevenLabs is now competing on this architecture, with a two-tier lineup that separates cost-sensitive conversation from higher-accuracy business tasks.

The OpenAI-compatible Realtime API lowers switching costs for teams already building voice agents, so adoption will depend on whether the latency and accuracy figures hold up in production. The headline numbers come from ElevenLabs and a single benchmark. Independent testing across noisy audio, accents, interruptions and tool calling would give a fuller picture.

The missing price figures, context limits and Mini benchmark results make cost-per-conversation comparisons impossible for now. Builders should wait for the pricing details and run their own evaluations before committing.