ARC-AGI-3 Announcement
Source: https://x.com (tweet) Author: ARC Prize (inferred) Date: 2025-2026
Summary
Announcement of ARC-AGI-3, described as “the only unsaturated agentic intelligence benchmark in the world.” Humans score 100%, AI scores <1%, demonstrating the benchmark specifically tests what current AI cannot do.
Key Claims
- ARC-AGI-3 is designed to test capabilities where humans score 100% and AI scores <1%
- Most benchmarks test what models have already learned to do
- The 99%+ human-AI gap is the stated evidence that AGI has not been achieved
- Specifically designed to resist saturation — tests genuine reasoning, not learned patterns
Concepts
- Reward Hacking — benchmark saturation as the problem ARC-AGI-3 tries to avoid
- Scaling & Compute — benchmark signals what scaling has not solved
Notes
Complements OpenAI’s real-world task evaluation — both motivated by benchmark saturation. ARC takes the opposite approach: find tasks AI provably cannot yet do.