The same model scored 62.7% and 99.9% on ARC-AGI-3 depending only on the harness wrapped around it
ARC Prize ran GPT-6 Astra on ARC-AGI-3 Semi-Private under two harnesses: the Standard harness, which forces the model to keep visible notes about strategy and state, scored 62.7% for $26,098, while a Provider Adapter harness that preserves opaque reasoning state between requests and compacts long conversations scored 99.9% for $18,817. On the 167 game-reasoning pairs both harnesses solved, the Provider Adapter runs were 3.66x faster and used 49% fewer total tokens. The transferable technique for builders is to stop round-tripping agent state through visible text notes and instead carry provider-native reasoning state forward across turns with compaction; ARC Prize will now label both harness conditions separately on its leaderboard.
Source
↳ Follow the thread