OpenAI's GPT-6 Astra on ARC-AGI-3

OpenAI's GPT-6, known as Astra, has recently achieved impressive results on the ARC-AGI-3 benchmark, which measures agentic intelligence in artificial intelligence systems. ARC-AGI-3 is the third generation of the ARC-AGI benchmark series, with the goal of measuring the "residual gap" between current AI and AGI (Artificial General Intelligence), defined as a system's ability to acquire any skill a human can, as efficiently as a human can.

In the Standard harness, Astra scored 62.7% on ARC-AGI-3, while in the Semi-Private for $26K configuration, it scored 99.9%. These scores represent state-of-the-art performance. Astra's replays demonstrate its ability to turn unfamiliar game mechanics into working models through three key findings: compact algebraic notation, action efficiency compared to humans, and custom tools it builds.

Astra develops compact algebraic notation, which enables it to track objects, coordinates, rules, and unfinished plans efficiently. It also uses a custom domain-specific language notation for generating these environment models. In the Provider Adapter harness, Astra's action efficiency exceeded the human baseline on 96.0% of levels, using 51.7% fewer actions on average. This milestone demonstrates that Astra can match and even surpass human parity in action efficiency, a critical measure of AGI.

Before ARC-AGI-3's launch, OpenAI tested approximately 500 participants to establish a human baseline for action efficiency. Astra's performance shows that it requires fewer interactions than the human baseline to execute solutions, indicating its efficiency in learning to solve new environments. The PRO-LONG harness, an ARC-AGI-3 red-teaming partner, further highlights Astra's capabilities with external tools, such as board parsers, game-state models, search algorithms, planners, and persistent notes.