Give a system a game without instructions. Can it discover what matters and plan its next move? In its 3 September 2026 evaluation, ARC Prize used that kind of challenge to study GPT-6 Astra. The numbers become much more useful once we look at how they were produced.
What the games measure
ARC-AGI-3 uses abstract, turn-based environments. An agent explores, infers rules and goals, and acts without being handed a set of instructions. The benchmark studies those skills inside a deliberately bounded setting.
Human solvability is easy to misread. It does not mean every participant solved every game. The human study says at least two people completed each environment. The score also considers action efficiency against a human reference, rather than completion alone.
One model, different conditions
On the semi-private set, ARC Prize reported Astra’s best score as 62.7% using its standard harness and 99.9% with a separate provider adapter. These best results used different reasoning settings. The adapter preserves more internal context between requests, so the two figures should be read alongside their different conditions.
In the adapter setup, Astra at maximum reasoning used fewer actions than the human baseline on 96% of levels, according to ARC Prize. That comparison concerns action efficiency in these games; it does not compare every human ability.
What the score cannot tell us
The environments have bounded, closed-ended mechanics. A strong result therefore does not by itself prove that a system can independently handle every unfamiliar real-world task or that artificial general intelligence has arrived. When a new score appears, ask three questions: which benchmark, which setup and which human reference?
