ChainThink report, March 26, 2026: According to 1M AI News monitoring, the ARC Prize Foundation, a non-profit institution co-founded by François Chollet, creator of Keras, and Mike Knoop, co-founder of Zapier, has launched the ARC-AGI-3 benchmark.

Differing from the static grid reasoning tasks of its predecessors, ARC-AGI-3 is an interactive turn-based environment where Agents operate within a 64×64, 16-color grid world. Without predefined instructions or goal hints, Agents must autonomously explore the environment, infer underlying rules and victory conditions, construct a world model, and plan action sequences. Scoring is based on "action efficiency"—fewer steps required to complete a level yield higher scores—distinguishing genuine reasoning capability from brute-force enumeration. All environments have undergone human calibration testing, confirming 100% solvability by humans upon first exposure.

As of launch, top-performing AI models achieved the following scores: Google Gemini 3.1 Pro Preview 0.37%, OpenAI GPT 5.4 (High) 0.26%, Anthropic Opus 4.6 (Max) 0.25%, xAI Grok-4.20 (Beta) 0.00%.

The release of this new version stems partly from concerns over contamination of prior benchmarks. Previously, Gemini 3 automatically leveraged the integer-to-color mapping intrinsic to ARC-AGI within its reasoning chains, despite no mention of such mapping in prompts—indicating that model training data had fully encompassed ARC-AGI tasks. ARC-AGI-3 resists such memorization shortcuts through interactive environments and autonomous goal discovery mechanisms. The ARC Prize 2026 competition features total prize funds exceeding $2 million.