OpenAI's GPT-6 Astra Achieves Near-Perfect Scores on ARC-AGI-3 Benchmark

OpenAI's GPT-6 Astra has achieved impressive results on the ARC-AGI-3 benchmark, a test designed to evaluate agentic intelligence through abstract, turn-based environments. Using two different evaluation setups, Astra scored 62.7% with the Standard harness and reached 99.9% with the Provider Adapter harness, both representing state-of-the-art performance. ARC-AGI-3 challenges AI agents to explore unfamiliar environments, infer goals, and build internal models to plan actions effectively without explicit instructions. The benchmark measures four key components: exploration, modeling, goal-setting, and planning and execution. Human participants can solve all test environments, providing a baseline for comparison. Astra's performance is notable not only for its high scores but also for its efficiency. It used fewer actions than the median human player on 96% of the levels, averaging 51.7% fewer actions per level. This surpasses the human baseline for action efficiency, marking a significant milestone in AI capabilities. One of Astra's distinguishing behaviors is its ability to create compact symbolic world models. The model generates a custom domain-specific language to represent game mechanics as logical rules, tracking objects, coordinates, and plans with precise, information-dense notation. This algebraic shorthand helps Astra plan and execute multi-step actions effectively. In an advanced evaluation setup called PRO-LONG, Astra demonstrated the ability to build custom tools such as board parsers, game-state models, and search algorithms. For example, in a maze-like game, it developed specialized software modules to handle navigation, combat, and patrol prediction, showcasing its capacity to extend reasoning through external code execution. The two harnesses used in testing highlight different aspects of model performance. The Standard harness provides a minimal interface requiring the model to manage its own notes and reasoning state, while the Provider Adapter harness allows preservation of opaque reasoning states and compaction for longer conversations. Astra's near-perfect score was achieved using the latter, which also improved efficiency by reducing token usage and elapsed time. While Astra's achievements represent meaningful progress toward artificial general intelligence (AGI), the ARC-AGI-3 benchmark itself is limited to deterministic, closed-ended environments and does not capture the full complexity of real-world scenarios. The ARC-AGI series continues to evolve alongside AI advancements, aiming to measure and challenge emerging capabilities. OpenAI's GPT-6 Astra thus marks a significant step forward in AI's ability to generalize, plan, and act efficiently in novel environments, providing valuable insights for future research and benchmark development.