Astra beats humans on the one ARC-AGI metric that was supposed to hold

The ARC Prize team ran GPT-6 Astra on ARC-AGI-3, their interactive game benchmark, and the interesting result is not the top-line score. With their standard harness Astra reached 62.7% on the semi-private set; with a "Provider Adapter" harness it reached 99.9%. What the team flagged instead was action efficiency, a metric they had expected to stay on the human side for a while.

Astra used fewer actions than the human baseline on 96.0% of levels, and 51.7% fewer actions per level on average. The team had hypothesized that even when a model could eventually solve a game, it would flail through far more moves than a person to get there. That gap closed. The model developed a compact algebraic notation to track game state, and when given a sandbox it wrote its own tools, including parsers and solvers tailored to specific games.

Cost tells its own story. Runs ranged from $17K to $49K depending on reasoning effort, against roughly $12.78 per attempted game for human testers. Oddly, higher reasoning settings sometimes cost less, because the model solved games in fewer total actions. The team is careful: this is not proof of AGI, and ARC-AGI-3 has a bounded scope.

Why it matters

If you evaluate agents, action efficiency was a cheap proxy for "the model doesn't really understand the task, it's just brute-forcing." That proxy just got less reliable. You now need to measure whether an agent reaches the goal well, not merely whether it wanders to get there.