OpenAI GPT-6 Astra Scores 62.7% on ARC-AGI-3

OpenAI's GPT-6 Astra has achieved state-of-the-art results on the ARC-AGI-3 benchmark, scoring 62.7% for $26K using a Standard harness and 99.9% for $19K with a Provider Adapter harness.

GPT-6 Astra surpasses human baseline in action efficiency. The model used fewer actions than the median tested human on 96% of ARC-AGI-3 levels. It turns unfamiliar environments into compact symbolic world models, representing game mechanics as logical rules and developing its own domain-specific language shorthand to track state and plan actions.

ARC-AGI-3 measures agentic intelligence components. The benchmark tests exploration, modeling, goal-setting, and planning and execution through novel, abstract, turn-based environments. Agents must actively obtain information, build generalizable models, identify target states with sparse rewards, and map paths while course correcting.

These environments contain only core knowledge priors and are calibrated through controlled testing with human participants, who can solve 100% of them. ARC-AGI-3 expands on prior generations as frontier AI capabilities advance. The series aims to measure the residual gap to AGI, defined as acquiring any human skill as efficiently as a human.

Results vary by reasoning effort and harness type. With the Standard harness, which carries forward notes chosen by the model, Astra (max) scores 62.7% for $26,098 while Astra (high) scores 54.8% for $40,705. Higher reasoning levels generally cost less because Astra solves games in fewer actions, reducing model calls and tokens.

Using the Provider Adapter harness, which preserves opaque reasoning state between requests and uses compaction, Astra (high) reaches 99.9% for $18,817 and Astra (max) reaches 98.6% for $17,332. Both harness configurations set new state-of-the-art marks on the semi-private leaderboard. Full results are available on the ARC Prize site.

Cost and efficiency improve at higher reasoning. At max effort, Astra requires fewer actions than at lower levels, lowering total cost relative to medium, low, or none configurations. The benchmark's goal remains tracking progress toward human-level efficiency in acquiring new skills.