A harness called Schema paired with Opus 4.8 and a second model, Fable 5, scored 99% on the RHAE metric of the ARC-AGI-3 Public set, according to a post from Haven Feng on X on July 16, 2026. The same harness reached 95.35% when run with GPT-5.6 Sol. Both numbers sit far above what either model manages unassisted, which puts the credit on the harness rather than the underlying model.
ARC-AGI-3, the abstract reasoning benchmark built to resist memorization, drops an agent into a game environment without instructions, rules, or any description of what the pixels represent. The agent has to work out the mechanics purely from interaction. Prior approaches lean on the model’s raw pattern recognition to guess at cause and effect, which tends to break down once a game introduces a rule the model has not implicitly seen in training.
Schema takes a different approach. Instead of asking a model to intuit a game’s logic, it has the agent write that logic as an executable program, essentially a working hypothesis of how the environment behaves. The program generates predictions, those predictions get checked against what actually happens on screen, and the agent revises the program when reality contradicts it. Planning then happens inside the corrected model rather than inside the raw pixel stream. Feng described the effect as making a language model reason like a physicist, building a falsifiable theory and updating it on contact with evidence.
The gap between the two model pairings is the more interesting data point for builders. Opus 4.8 with Fable 5 clears 99% RHAE; GPT-5.6 Sol lands at 95.35% under the identical harness. A roughly four-point spread on the same scaffolding suggests the harness supplies most of the reasoning structure, while the base model still determines how reliably it writes and debugs the executable hypotheses Schema depends on. That is a meaningfully different failure mode than a model simply guessing wrong.
The result also reframes what “scaling” buys on abstract-reasoning tasks. ARC-AGI-3 was designed specifically to punish memorization and reward genuine world-modeling, and frontier labs have historically posted modest scores on it even as they climbed other leaderboards. A harness that converts perception into code, tests that code, and plans against the tested version is a structural fix rather than a bigger model. It treats the benchmark’s core demand, understanding an unexplained system, as a software-engineering problem instead of a prediction problem.
Feng’s post is a company announcement, not an independent evaluation, and it does not include third-party replication of the RHAE figures or a comparison against other ARC-AGI-3 harnesses attempted by other labs. Readers should treat 99% and 95.35% as self-reported until the ARC Prize organization or another outside party verifies the runs.
For teams building agents meant to operate in genuinely novel environments, such as new codebases, unfamiliar APIs, or physical or simulated systems without documentation, the takeaway is architectural: pairing a capable model with a harness that forces it to write and test explicit hypotheses may close more of the reasoning gap than swapping in a larger model. Anyone evaluating agent frameworks this quarter should ask whether the framework tests its own predictions against ground truth, not just whether it uses a stronger base model.
Haven Feng announced Schema’s ARC-AGI-3 results in a post on X on July 16, 2026.
