The ARC Prize has launched the ARC-AGI-2 benchmark as part of its 2025 competition, offering a total of $1 million in prizes to drive advancements in artificial general intelligence (AGI).

The grand prize of $700,000 will be awarded for achieving 85% success within specified efficiency limits, alongside additional categories for top scores and transformative research ideas.

This competition will take place on Kaggle, featuring a live leaderboard and various prize categories to incentivize innovative solutions to the ARC-AGI-2 challenges.

The ARC-AGI-2 benchmark aims to identify capability gaps in AI and guide innovation toward achieving general adaptive intelligence, moving beyond simple memorization tasks.

Designed to be challenging for AI but relatively easy for humans, ARC-AGI-2 emphasizes tasks that are difficult or impossible for AI to solve, highlighting the adaptability that characterizes human intelligence.

The benchmark focuses on symbolic interpretation, compositional reasoning, and contextual rule application, areas where AI typically struggles compared to human performance.

Efficiency in problem-solving is now viewed as a critical factor in determining intelligence, with metrics showing significant disparities between human and AI effectiveness.

In a practical example, human panels achieve 100% accuracy on ARC-AGI-2 tasks at $17 per task, while OpenAI's o3 only achieves 4% accuracy at a cost of $200 per task.

The introduction of OpenAI’s o3 in late 2024 marked a significant advancement in AI, showcasing a transition from rote memorization to more sophisticated reasoning capabilities, though still requiring substantial human oversight.

Established in 2019, the ARC Prize has been pivotal in creating benchmarks that encourage researchers to push towards AGI by measuring fluid intelligence and inspiring new ideas.