Keeping the Model Intact, Harness Change Yields 99% Human Efficiency Near ARC-AGI-3

A case demonstrating that the core of AI inference performance improvement lies in harness design rather than the model itself has emerged. Impossible Research released a new agent execution framework called Schema, achieving 98.98% performance on the ARC-AGI-3 benchmark. The researchers emphasized that this result shows execution method enhancements without altering model weights can significantly boost AI performance, positioning agent architecture and execution frameworks as critical competitive advantages. Schema achieved 98.98% RHAE with Opus 4.8 + Fable 5 and 95.35% with GPT-5.6 Sol on the ARC-AGI-3 public set, surpassing previous records. However, the results are self-measured and lack independent ARC Prize verification.