ARC Prize has introduced ARC-AGI-2, its most demanding AI benchmark to date, alongside the launch of its 2025 competition, offering $1 million in prizes. Designed to highlight AI’s adaptability gaps, ARC-AGI-2 challenges systems beyond mere memorization, encouraging genuine reasoning and problem-solving skills.
Since its 2019 inception, the ARC Prize has provided AGI researchers with benchmarks that measure and inspire progress. The ARC-AGI-1 benchmark, for instance, tested fluid intelligence, pushing AI beyond rote learning.
The arrival of OpenAI’s o3 in late 2024, which combined large language models with reasoning synthesis, marked a breakthrough in AI capabilities. However, such systems remain inefficient and require extensive human supervision. ARC-AGI-2 aims to challenge AI further, ensuring progress towards adaptable and resource-efficient intelligence.
Unlike many AI benchmarks that test superhuman abilities, ARC-AGI-2 focuses on tasks easy for humans but difficult for AI. The benchmark assesses symbolic interpretation, compositional reasoning, and contextual rule application—areas where AI still struggles. Humans solve every ARC-AGI-2 task in under two attempts, while top AI models achieve only single-digit success rates.
A key factor in achieving AGI is efficiency. ARC Prize highlights cost per task as a crucial metric, revealing stark differences between humans and AI. Humans complete ARC-AGI-2 tasks with 100% accuracy at $17 per task, whereas OpenAI’s o3 has an estimated 4% success rate at $200 per task. This efficiency gap underscores AI’s limitations in adaptive problem-solving.
The ARC Prize 2025 competition, hosted on Kaggle, will award $1 million across categories, including a $700,000 grand prize for reaching 85% success within efficiency constraints. Additional prizes will reward top scores and innovative research contributions.
With increased stakes, the ARC Prize aims to drive AI beyond scaling existing models, encouraging groundbreaking ideas that redefine AI efficiency and adaptability.
