François Chollet: ARC-AGI-3 Launching March 25 — First Interactive Reasoning Benchmark Requires Agents to Explore, Learn, Plan, and Adapt

ARC Prize·medium signal

The ARC Prize Foundation will release ARC-AGI-3 on March 25, fundamentally redesigning what AI reasoning benchmarks measure: 1,000+ levels across 150+ environments that require agents to explore unknown contexts, form hypotheses, and adapt based on feedback rather than match static patterns. This directly targets the generalization gap that all current frontier models — including those scoring 80%+ on SWE-bench — have failed to close, since ARC-AGI-3 cannot be gamed through memorization or compute scaling. The launch follows ARC Prize 2025 closing with 1,454 teams and no solution clearing the benchmark bar.