TL;DR: A new inference-time harness called Schema reaches 99% on the ARC-AGI-3 Public set by wrapping Claude Opus 4.8 and Fable 5 without modifying model weights, demonstrating that process-level improvements can dramatically boost benchmark performance.
Summary: The Schema harness, announced by an independent team, achieves 99% on the ARC-AGI-3 Public set using Claude Opus 4.8 and Fable 5, and 95.35% with GPT-5.6 Sol. It does not alter model weights but instead changes how observations are converted into game models, how predictions are tested against interaction history, and how plans are executed and revised. A fallback rule reruns low-scoring games with stronger models to retain the higher per-game score.
Why it matters: For AI builders, Schema shows that significant benchmark gains can come from smarter orchestration and inference strategies rather than new model architectures—this opens up a design space for improving performance on reasoning tasks without retraining. Consider evaluating Schema's approach for your own evaluation pipelines or experimenting with its process-level improvements on similar benchmarks.
Source: reddit