Home › Benchmarks › ARC-AGI
ARC-AGI
Benchmark
Capability
Chollet's abstraction puzzles — designed to resist everything but intelligence.
You're looking at a benchmark record: a test the industry uses to measure chips or models — and what it actually tells you.
ARC's colored-grid puzzles require inferring a novel rule from a few examples — trivial for humans, brutal for pattern-matchers. o3's 87.5% high-compute result on ARC-AGI-1 (Dec 2024) was a landmark; ARC-AGI-2 promptly reset the frontier to single digits, keeping the target moving and the definition debate alive.
How this connects to the rest of the stack
This record's links at a glance — click a group to jump to it.
Benchmarks
Sources
Record reviewed as of 2026-01-15 — see Methodology.