Create an AI capable of novel reasoning.
Real intelligence isn’t about memorizing answers - it’s knowing what to do when the problem changes.
Build systems that learn and adapt to novel, human-solvable tasks they’ve never seen before and advance AI’s ability to learn new skills efficiently.
Note: This is a relaunch of ARC Prize 2025. If you are joining either this or ARC-AGI-3, please consider joining the paper track, where you can document your approach for either one of the prediction competitions.
Today’s AI systems excel at what they were trained to do, but often fall short when something unfamiliar comes along. Most benchmarks reward pattern recognition, not genuine problem-solving. The ARC Prize focuses on true generalization: whether a system can quickly learn new skills in unfamiliar situations. Instead of rewarding pattern recognition on known tasks, it evaluates how well systems can adapt to new problems they’ve never encountered before. The evaluation environment is designed so systems can't just memorize solutions. Tasks take place in hidden, interactive environments that require exploration and multi-step reasoning.
This competition evaluates submissions on the percentage of correct predictions. For each task, you should make 2 attempts to predict the exact outputs for every test input grid contained in the task. The final score is the sum averaged of the highest score per task output divided by the total number of task test outputs.
The submission file for this competition must be a json named submission.json. For each task output in the evaluation set, you should make exactly 2 predictions (attempt_1, attempt_2). All the task_ids in the input challenges json file must also be present in the submission.json file.
TOTAL PRIZES AVAILABLE: $700,000 - Progress Prizes: $275,000 - Grand Prize: $275,000 - Bonus Prize: $150,000
In line with the spirit of the competition, participants eligible for a prize will be removed from the competition if they do not open source their solutions. A Bonus Prize of an additional $150,000 will be unlocked in the event that a team achieves a score of at least 85% accuracy on the competition leaderboard.
This is a Code Competition. Submissions to this competition must be made through Notebooks. In order for the "Submit to Competition" button to be active after a commit, the following conditions must be met:
- CPU Notebook <= 12 hours run-time
- GPU Notebook <= 12 hours run-time
- No internet access enabled
- External data, freely & publicly available, is allowed, including pre-trained models
- Submission file must be named submission.json