Two harnesses put the same model 37 points apart on ARC-AGI-3

The two conditions differ in what the model is permitted to remember. ARC Prize's standard harness gives Astra continuity only through notes it writes to itself and chooses to carry forward, so anything it does not write down is absent on the next request ^[2]. The provider adapter harness preserves the model's opaque reasoning state between requests and compacts the conversation as it lengthens, letting prior work be reused rather than rebuilt ^[4]. The distance between the two results is 37.2 points, 99.9 minus 62.7 ^[16].

Suppose most of that spread belongs to the harness, which is the reading the write-up invites. Then 37.2 points is an estimate of how much a frontier model loses when it has to compress its own working state into prose for its future self. The replays show what survived that compression: object positions, coordinates, rules, unfinished plans, and a shorthand Astra generated per environment, with lines such as "extend8 to3; retract10 to2; shorten8 to1" standing in for a multi-step plan ^[8]. Dense self-summarization, and on this evidence still lossy against whatever the provider keeps internally ^[4].

The size of the spread is what complicates the board. Astra's standard-harness 62.7% sits 32.5 points above the 30.2% ARC Prize reports for Claude Opus 5 ^[17], which itself sits well above GPT-5.6 Sol at 7.8% ^[10]. So the gap between one model's two entries is wider than the gap between the top two systems ^[18].

The human costs in the same write-up are a lesson in denominators. Participants earned $115 for a 90-minute session plus $5 per completed game and attempted about nine games, roughly $12.78 per attempted game before bonuses ^[11]. ARC Prize also prices the brain's electricity alone at about 0.067 cents per game attempted, noting that most of the fee buys a participant's time and willingness rather than their energy ^[12]. Those two ways of costing identical human performance differ by a factor near 19,000 ^[19]. The agent side has no comparable per-game figure at all: both dollar totals cover a whole benchmark run, and no task count is published beside them ^[14].

The action-efficiency result is narrower than it first sounds. Astra used fewer actions than the median tested human on 96% of levels ^[5], which counts moves spent, not levels finished. And the thing this doesn't tell you is how far these environments carry into the agent loop you actually run: they are turn-based, withhold instructions, and require the agent to infer its own goals ^[7], which is the setting where uninterrupted private reasoning should pay most. With the leader now a tenth of a point from the ceiling ^[20] and humans solving every environment ^[6], ARC-AGI-3 has little room left to distinguish a good agent from a good harness.