Astra hits 62.7% on ARC-AGI-3 — and 99.9% with a catch

OpenAI's GPT-6 Astra scored 62.7% on the ARC-AGI-3 semi-private set under ARC Prize's neutral Standard harness, at a cost of roughly $26,000. Swap in OpenAI's own Provider Adapter harness — which keeps opaque reasoning state between requests and compacts long contexts — and the same model hits 99.9% for $19K. Same weights, wildly different number.

The score isn't the interesting part. Astra beat the human action-efficiency baseline: fewer moves than the median tested human on 96% of levels, and 51.7% fewer actions per level on average. It got there by compressing unfamiliar games into compact symbolic world models — terse notes like "L8: hub q2 (8↓)" tracking objects, coordinates and interaction rules in an algebraic shorthand it invented on the fly. Not quite a program, but close.

ARC Prize's framing is careful: publish both numbers, label the evaluation condition every time. Human testers, for reference, averaged around 48% on the same games.

Why it matters: the harness is now part of the model — a benchmark number without its eval config is marketing, not a measurement.