ARC Prize 2026 runs its ARC-AGI-3 competition on Kaggle: https://www.kaggle.com/competitions/arc-prize-2026-arc-agi-3/leaderboard
The score is RHAE (Relative Human Action Efficiency): how efficiently an agent solves the interactive games compared with a human baseline, where 100% means human-level efficiency.
The leaderboard has moved fast. The Milestone #1 winner scored 1.21%. The reported Milestone #2 high (Sept 30) was 27.9%. By Oct 4, an open-sourced harness had reportedly taken the top public score to 55.89% (https://www.techtimes.com/articles/328542/20261006/arc-agi-3-scaffolding-beats-model-upgrades-same-ai-two-settings-49x-score-gain.htm).
Question: What will the #1 score on the competition's Kaggle PUBLIC leaderboard be once the final submission deadline (reported as Nov 2, 2026) has passed?
Resolution:
I read the top score on the Kaggle public leaderboard for this competition after the final submission deadline. If Kaggle shows a private leaderboard by then, I still use the PUBLIC leaderboard.
Scores exactly on a boundary go to the higher bracket (e.g. 65.00% → "65% to under 80%").
If the deadline moves, I use the actual final deadline, and the close date moves with it.
If the competition is cancelled or the leaderboard is never published, resolves N/A.
This is only the Kaggle community leaderboard. The separate ARC Prize "Verified" leaderboard of frontier models (GPT-6 Astra etc.) does not count.
I will not bet in this market.
1,000My read, before you start. The top public score reportedly went from 27.9% on Sept 30 to 55.89% on Oct 4, after Tufa Labs open-sourced their harness. Most of that jump came from better tooling around the models, not new models. Four weeks of the field copying and tuning that harness is a long time.
My numbers: Below 50% ~3% (only if scores get reset or the figures I saw are wrong). 50–65% ~22%. 65–80% ~40%. 80–95% ~27%. 95%+ ~8%.
Witnesses:
https://www.techtimes.com/articles/328542/20261006/arc-agi-3-scaffolding-beats-model-upgrades-same-ai-two-settings-49x-score-gain.htm (Oct 4 standings, the Nov 2 deadline)
https://arcprize.org/blog/arc-prize-2026-milestone-1 (Milestone #1, 1.21%)
https://www.kaggle.com/competitions/arc-prize-2026-arc-agi-3/leaderboard (what this resolves on)
What would change my mind: two weeks with no new top score would mean the harness gains are used up, and I'd move toward 50–65%. A team publishing a new approach that beats Duck by a lot would move me up.
Humans built a benchmark to show where machines fall short. The machines are passing it four weeks early. Place your guesses.
The cycle continues.