Safety, Ethics & Governance · advanced · concept 293 of 311 · Updated Nov 4, 2026
ARC-AGI and Reasoning Benchmarks
ARC-AGI is a test of fluid intelligence made of small grid puzzles that people solve on sight but that cannot be answered from memorized data. François Chollet introduced it in 2019 to measure how well a system acquires a new skill, not how many skills it has stored. It resists the saturation that hit older benchmarks. In December 2024 OpenAI's o3 scored 75.7% on the semi-private set within the compute limit; ARC-AGI-2 followed in March 2025.
Top AI contributors on this
- Demis HassabisGoogle DeepMind (chair) / Isomorphic Labs
- Ilya SutskeverSafe Superintelligence Inc. (SSI)
- Sam AltmanOpenAI
- David SilverIneffable Intelligence / UCL
- Elon MuskTesla / SpaceX
- Eliezer YudkowskyMachine Intelligence Research Institute (MIRI)
Key terms
Learn these first
Where you meet it in the real world
Tracking frontier model reasoning progress, ARC Prize Kaggle entries, lab capability claims and safety reporting, choosing models for novel-problem work
Videos
ARC Prize · YouTube
AI Explained · YouTube
Lex Fridman · YouTube
Guides and articles
Courses, papers, and more