It's remarkable how well GPT-6 Astra did on ARC-AGI-3 even when its reasoning effort was set to "none". I highlighted the two "none" data points in red in the leaderboard chart below to make them easy to see. Note that "none" was available during early testing but is not currently available via the API. In the Standard harness, which lets the model decide which notes to carry forward, Astra scored 35.2% using "none", beating its 17.5% score at "low" and coming close to its 38.6% score at…