Atlas

Benchmarks

← All benchmarks

InteractiveBench Accessibility (task pass)

Agents · 2026-10-25

InteractiveBench's Accessibility category: the 20 Accessibility tasks, in which the Human is a blind or low-vision user who receives no frames, so the assistant must describe the scene and guide every action (for example cleaning a kitchen with broken glass). Same simulator, Gemini 3.1 Pro Human agent and end-state check as the full benchmark; the score is the percentage of these 20 tasks in which every required end state holds.

Top models (higher is better)

ModelScore
GLM-5.3 Flash50.0
GPT-5.545.0
Kimi K340.0
Gemini 3.1 Pro Preview35.0
Muse Spark 1.135.0
Claude Fable 5.130.0
Opus 4.830.0
GPT-5.6 Sol30.0
Claude Opus 525.0
GPT-5.6 Luna20.0
Loading Atlas data…