OpenClaw is now sitting at #1 on the official ARC-AGI-3 community leaderboard — the benchmark designed by François Chollet's ARC Prize Foundation to measure genuine fluid intelligence in AI systems. The submission, listed simply as "OpenClaw," scored 5.2% on the benchmark at a total compute cost of $2,912, with a run date of May 15, 2026.
What ARC-AGI-3 Is
ARC-AGI (Abstraction and Reasoning Corpus) is widely considered one of the harder tests of AI reasoning that can't be easily gamed by scale alone. The third generation — ARC-AGI-3 — pushed the difficulty further, and the human baseline sits at 95.3%, which puts the 5.2% OpenClaw result in context: it's a first position on the community board, but there's a long road ahead.
That said, landing the top community slot with an agent harness rather than a purpose-built model is notable. The entry description reads:
"OpenClaw Harness adapted to play ARC-AGI-3 allowed memory and code execution tools."
The ARC Prize Foundation published the full code and a scorecard alongside the leaderboard entry.
Why This Matters
ARC-AGI-3 submissions from the community have historically come from purpose-built systems — evolutionary search algorithms, fine-tuned models, or domain-specific solvers. Seeing an out-of-the-box agent harness like OpenClaw compete on the same benchmark, using memory and code execution as its primary tools, signals something interesting about where general-purpose agents are heading.
The submission also validates OpenClaw's long-held design philosophy: a capable agent shouldn't need task-specific scaffolding to reason about novel problems. It should be able to use the tools it already has — code execution, memory — and get somewhere.
The Competition
The next entries on the ARC-AGI-3 board include a "Read-Grep-Bash Agent" from researchers at Duke and other institutions, whose score is still pending a full public-set run. Below that, the leaderboard shifts to ARC-AGI-2 entries from 2025.
Contribute or Replicate
The agent template is open source under the ARC Prize Foundation's repository. If you want to run your own OpenClaw agent against ARC-AGI-3, the template is a reasonable starting point. The benchmark itself is available at arcprize.org.