Schema: a harness for llms, with Fable+4.8 or GPT 5.6 Sol, (supposedly) achieves 99% and 95.35% respectively on ARC-AGI-3.
https://schema-harness.github.io/
Edit: IMPORTANT CLARIFICATION
This is just the public set, which is much easier than the held out set, "supposedly" is the wrong way to frame it, they seemingly did on the public set, but what would be interesting is to see how it performs on the held out set.