July 16, 2026
99%… and 100% side-eye
Schema Harness Achieves ~99% on Arc‑AGI‑3 Public
AI scores 99% on a brutal puzzle test, but the comments are screaming “show us the real score”
TLDR: Schema got an AI system to nearly ace a hard puzzle benchmark without changing the underlying model, which is a big deal because it suggests smarter setup can matter as much as smarter AI. The comments, though, were dominated by skepticism: people want private test results and proof this isn’t just a benchmark party trick.
A new AI setup called Schema just posted a jaw-dropping ~99% on the public ARC-AGI-3 test, a benchmark built around weird little games that don’t explain the rules. In plain English: the AI gets a colorful grid, a few buttons it can press, and basically has to figure out what on earth is going on by trial and error. That alone is headline-worthy. But in the comments, the crowd immediately turned this victory lap into a courtroom drama.
The biggest reaction? Suspicion, curiosity, and a lot of “okay, but…” energy. Multiple commenters zeroed in on the same question: if this is such a breakthrough, what happened on the private test set? That became the thread’s unofficial chant. One person bluntly asked whether this really generalizes to other games like Atari or Nintendo, or whether it’s just sneaking in game-specific hints—the exact kind of benchmark gaming critics have warned about for years. Another commenter cut through the hype with a wonderfully basic question: what does 99% even mean here? Honestly, relatable.
Still, the optimism squad showed up too. Some saw possible ripple effects for coding assistants and broader AI progress if this method really helps models reason better without changing the model itself. And in classic internet fashion, someone brought receipts from a Baba Is You test, basically saying: yes, the puzzle-game arms race is getting real. The vibe was equal parts breakthrough buzz, trust issues, and nerdy excitement—with everyone agreeing on one thing: this result is flashy, but the real tea is whether it survives harder tests.
Key Points
- •ARC-AGI-3 gives agents only a 64×64 color-grid observation and legal actions, without explicit rules, goals, object lists, or shaped rewards.
- •The benchmark’s official metric is Relative Human Action Efficiency (RHAE), where 100% means completing every level at or above human-baseline action efficiency.
- •The article reports frontier-model progress on the semi-private set from 0.51% at launch in March to 7.78% in July with GPT-5.6 Sol at max reasoning, with Sol also scoring 13.33% on the public set.
- •Schema is presented as a harness that reaches 99% on the ARC-AGI-3 Public set with Claude Opus 4.8 and Fable 5, and 95.35% with GPT-5.6 Sol, without changing model weights.
- •Schema’s technical approach centers on jointly solving state grounding and mechanism discovery by encoding both state representation and transition rules in a single editable program.