ARC Prize measures GPT-6 Astra at 62.7 percent on ARC-AGI-3 under its neutral harness and 99.9 percent under the provider adapter

Sources: primary, discussion
Summary: ARC Prize reports a standard harness maximum of 62.7 percent at $26,098 and a Provider Adapter high setting of 99.9 percent at $18,817, so the provider-adapted run scored higher and cost less. ARC Prize states it will report both conditions on the leaderboard going forward and that it is not claiming AGI. Separate PRO-LONG runs gave the model a code sandbox that human test participants did not have.
Why it matters: At matched maximum effort the same model on the same benchmark moves 35.9 points, 98.6 percent through the provider adapter against 62.7 percent in the neutral harness, so a single headline score does not describe what the model will do inside a given harness.
Follow-up: Check whether the leaderboard publishes both harness conditions for later models.