On October 9, ARC Prize published the 2026 ARC-AGI-2 leaderboard, with TUFA Labs in first place at 88.06%. Second through fifth are Rabbithole (80.56%), Yi-Chia Chen (77.22%), Nubanana (77.08%) and _hans (67.64%). The same day, ARC Prize co-founder François Chollet also noted publicly that the ARC-AGI-2 score on Kaggle had reached 88.06%.
The leader is about 7.5 percentage points ahead of second place, and among the top five only TUFA Labs has cleared 85%.
The 85% line
ARC-AGI-2 is an abstract reasoning test designed by Chollet and colleagues. Each task shows a few input-output examples built from colored grids; the system has to infer the underlying rule and then draw the correct grid for a new input. Each test input allows 2 attempts, and a task scores 1 point if either attempt is exactly right. The tasks deliberately avoid anything that can be solved by recalling memorized knowledge.
The contest runs on Kaggle, in an offline evaluation environment with a compute cap. According to the 2026 rules page, the total prize pool is $700,000:
- $275,000 for the Progress Prize, split among the top eight places, with $75,000 going to first place;
- $275,000 for the Grand Prize, awarded to the highest-scoring write-up of a solution;
- a further $150,000, conditional on reaching 85% on the private evaluation set.
All winning solutions must be open-sourced, and entries that are not will be removed from the competition.
One detail needs to be read carefully. ARC Prize's announcement says the $150,000 is shared by all teams that exceed 85%, while the rules page says it goes to the "first" qualifying solution to reach that mark on the private set. The organizers have not yet explained which wording takes precedence.
Also, the public Kaggle leaderboard is scored on only about half of the test questions, and the final ranking is decided by the other half. The 88.06% is an interim result, and whether TUFA Labs can hold 85% on the private set will not be known until the season is settled.
Putting the two seasons side by side
ARC Prize 2025 also used ARC-AGI-2. That year 1,455 teams made 15,154 submissions, and the winner, NVARC, scored 24.03% on the private set at roughly $0.20 per task. Runner-up the ARChitects scored 16.53%, and nobody claimed the Grand Prize. NVARC is a team of two NVIDIA Kaggle grandmasters whose approach combined synthetic data, test-time training and a small model, rather than brute-forcing the problems with a large one.
Going back further, the 2024 season used the original ARC-AGI, and the winning score was a little over 50%. When ARC-AGI-2 launched in 2025, its difficulty was deliberately raised, and leading frontier models generally scored only in the single digits on it at the time.
So the best Kaggle result has gone from 24% to 88% in a year. The two numbers are not measured the same way: the former is a final private-set score and the latter is an interim public-leaderboard score, so subtracting one from the other would overstate the improvement. Even so, the change in order of magnitude is large.
TUFA Labs has not publicly described its method, and neither the compute consumed nor the cost per task behind the result has been disclosed. Under the rules, winning solutions must be open-sourced and submitted with a write-up, so details will not be available until the season ends at the earliest.
How much room is left in the tasks
When the organizers released ARC-AGI-2 in 2025, they published a human baseline: test takers averaged about 60% correct, and every task was solved by at least two of them. On that basis, a public-leaderboard score of 88% is already above the average human test taker.
ARC Prize has also been shifting its focus over the past two years toward the interactive ARC-AGI-3, whose tasks are small games that require learning by trying. If the private set confirms a score above 85% on ARC-AGI-2 this season, this static grid format will most likely have run its course as a prize-money benchmark. The organizers have not yet announced what comes next.
Sources: ARC Prize official account, ARC Prize 2026 competition rules page, CocoLoop, ARC Prize 2025 results analysis; used to verify the top-five scores, the prize structure and the basis of the 2025 winning score.