Schema harness hits 99% on ARC-AGI-3 without touching model weights
A new harness called Schema — an external framework that wraps around an AI model to control how it's used — reportedly reaches 99% on the ARC-AGI-3 public benchmark when paired with and Fable 5, and 95.35% when paired with GPT-5.6 Sol. Crucially, it doesn't modify the underlying (the model's internal learned parameters) at all. Instead, it changes the process wrapped around the model: how raw observations get turned into a working model of the game's rules, how predictions get checked against the actual interaction history, and how plans get carried out and revised as new information comes in.
Scoring follows a fixed fallback rule: Opus 4.8 and Sol run first at their highest (xhigh); any game scoring below 80 gets rerun using Fable 5 or Sol at maximum reasoning (max), and the higher of the two scores is kept. The president of ARC Prize reacted publicly, saying it 'looks cool' and that they need to dig into it further.
Key points
- Schema harness scores 99% on ARC-AGI-3 public set with + Fable 5, and 95.35% with GPT-5.6 Sol
- are unchanged — only the surrounding process (observe, predict, verify, plan/execute) is different
- Games scoring below 80 get rerun at higher (Fable 5 or Sol max) and the better score is kept
- ARC Prize's president publicly noted interest and said the claim needs further investigation