Report
KNOWS leaderboard lists a 3.6% full-task success rate for its top AI agents
The benchmark’s creators say its 110 tasks require agents to turn live-web research into Google Docs, Sheets or Slides.
TLDR
The KNOWS benchmark team says GPT-6 Astra and GPT-6 Luna, both run through Codex, tied for the highest full-task success rate at 3.6% across its 110 browser-based tasks. Agents must research the live web and produce finished Google Docs, Sheets or Slides. The team says no evaluated system fully completed a Sheets or Slides task.
Combined views
1 Source, first seen ago
