ARC-AGI-3 Breaks Every Frontier Model: GPT-5.4 Scores 0.26%, Opus 4.6 0.25%, Grok-4.20 0.00% — Humans Score 100%

The Decoder·high signal

Chollet's ARC-AGI-3, launched March 25 at YC HQ with Sam Altman, is the first fully-interactive benchmark featuring hundreds of game-style environments with no instructions, no rules, and no stated goals. Gemini 3.1 Pro Preview led frontier models at just 0.37%, while simple CNN and graph-search approaches hit 12.58% — over 30x better than any LLM. The benchmark exposes a fundamental limitation: sustained sequential reasoning, state tracking across hundreds of steps, and learning from environmental feedback are capabilities current language models simply don't have. ARC Prize 2026 offers $2M+ in prizes, all solutions must be open-sourced.