ARC Prize 2026 - ARC-AGI-2
Create an AI capable of novel reasoning. Real intelligence isn't about memorizing answers—it's knowing what to do when the problem changes.
This competition challenges participants to build systems that learn and adapt to novel, human-solvable tasks they've never seen before, advancing AI's ability to learn new skills efficiently. Today's AI systems often excel at pattern recognition on known tasks but fall short with unfamiliar inputs. The ARC Prize focuses on true generalization: evaluating whether a system can quickly learn new skills in unfamiliar situations. The evaluation environment is designed so systems cannot simply memorize solutions; instead, tasks take place in hidden, interactive environments requiring exploration and multi-step reasoning.
Evaluation
Submissions are evaluated on the percentage of correct predictions. For each task, participants must make 2 attempts to predict the exact outputs for every test input grid. If either of the 2 predicted outputs matches the ground truth exactly, the submission scores 1 for that task output; otherwise, it scores 0. The final score is the sum averaged of the highest score per task output divided by the total number of task test outputs.
Timeline - Start Date: March 25, 2026 - Entry Deadline: October 26, 2026 (Participants must accept competition rules by this date) - Team Merger Deadline: October 26, 2026 - Final Submission Deadline: November 2, 2026 - Winners Announcement: December 4, 2026
Prizes A total of $700,000 is available: - Progress Prizes ($275,000): Distributed among the top 8 performers based on leaderboard accuracy. - Grand Prize ($275,000): Awarded to the highest scoring Solution Writeup. Submissions must be open-sourced and attached to an official writeup within seven days of the deadline. Evaluation criteria include Accuracy, Universality, Progress, Theory, Completeness, and Novelty. - Bonus Prize ($150,000): Unlocked if any team achieves at least 85% accuracy on the leaderboard. This prize is divided among the Top 5 qualifying teams.
Code Requirements
This is a Code Competition where submissions must be made through Notebooks. Key constraints include:
- CPU/GPU Notebook run-time must be ≤ 12 hours.
- No internet access is enabled during submission.
- External data (freely & publicly available) and pre-trained models are allowed.
- The submission file must be named submission.json.
Participants eligible for prizes must open-source their solutions. The competition offers access to Kaggle's L4x4 machines (96GB GPU memory), though these consume GPU quota at twice the rate of older machines and require internet to be disabled.