ARC Prize 2026: TUFA Labs posts 88.06% on ARC-AGI-2 under Kaggle compute limits, past the 85% bonus line, pending private evaluation

Why it matters

【Key points】

On Oct 9 (15:19 UTC), ARC Prize announced a new ARC-AGI-2 high score in ARC Prize 2026: 88.06% by TUFA Labs. Minutes earlier it said Yi-Chia Chen had taken the ARC-AGI-3 lead with 59.17%.

The track's goal is 85% on the ARC-AGI-2 private evaluation set within Kaggle efficiency limits: Kaggle notebooks, no internet during evaluation, and open-sourced code to be prize-eligible. Of the $700,000 in track prizes, a $150,000 bonus is reserved for solutions reaching 85%. Co-founder Mike Knoop said the threshold "just got hit on Kaggle" and that this is the final year of ARC-AGI-2 on Kaggle.

【Verifiable】

ARC Prize post: https://x.com/arcprize/status/2108578092558250148

ARC-AGI-3 post: https://x.com/arcprize/status/2108577347192676612

Mike Knoop: https://x.com/mikeknoop/status/2108584791436321097

Rules and prizes: https://arcprize.org/competitions/2026/arc-agi-2

Submissions close Nov 2 and results are announced Dec 4. Prizes are judged on the private set, and teams must open-source before receiving private scores, so 88.06% is a leaderboard score, not a final result.

【Our read】

1. The constraints are the story. Frontier systems post big ARC numbers with heavy compute and closed models; clearing 85% offline, compute-capped and open-source suggests these puzzles now yield to cheap, reproducible methods.

2. ARC-AGI-2 is not "solved" yet. Leaderboard and private scores can differ, and the bonus terms (first eligible solution on the site, split among teams in the post) await the December results.

3. Watch whether TUFA's open release credits test-time search or a stronger base model, and ARC-AGI-3, where 59.17% leaves real headroom.