ARC Prize 2026 · ARC-AGI-3
All articles
An open-weights agent on a single GPU, September–November 2026
ARC-AGI-3 is a benchmark of interactive puzzle games that an agent has never seen: no instructions, no stated goal, a small coloured grid and a handful of moves. These articles describe how we approached it with one open model on one GPU, what we measured, and what we learned.
Research notes: a fixed model on one GPU — what moves the score
The main lessons of the season, without scores and configurations until the competition's code deadline.
The engineering behind an ARC-AGI-3 solution
How one GPU turns model weights into tokens, and why that shapes what a competitive solution looks like.
When the whole team is AI
Notes from a project lead: what three months of managing AI agents taught me — seven failures, the rule each one produced, and the one thing rules could not fix.
Full write-up with all measurements
Every experiment with its numbers, the instruments we built, and their measured noise.
The road travelled and the lessons learned
Two months on ARC-AGI-3: the score, the engineering, the measurement tools, and the lessons that cost the most.