ARC-AGI-2 launches; frontier models score in single digits

The harder successor to ARC-AGI sees pure LLMs at 0% and reasoning models at only a few percent, versus ~60% for human panels.

What happened and why it matters

ARC-AGI-2 was designed after o3 to remove brute-force-searchable tasks. The results showed that the efficient, flexible generalisation of humans had not yet been matched.

Sources

• Announcing ARC-AGI-2 and ARC Prize 2025 ↗ ARC Prize · 24 Mar 2025