A new system called Schema has achieved approximately 99% performance on the ARC-AGI-3 benchmark, a challenging task where agents must infer game rules from raw visual input without explicit instructions.

Schema utilizes Claude Opus 4.8 and Fable 5 to reach this score, demonstrating a significant improvement over previous frontier models. The system's success lies not in altering underlying model weights, but in refining the process of how observations are used to build and test game models. It formalizes key problems: state grounding (identifying objects and properties) and mechanism discovery (understanding how actions change the state), solving them jointly.

This approach mirrors how physicists develop theories, by jointly refining their understanding of the state and the laws governing it. Schema's latent world representation is a program, making it interpretable and verifiable. The system's control loop involves observing, deliberating, executing actions, and recording outcomes, allowing for continuous refinement of its game model and plans.