FrontierCode 2.0: Astra trails Opus 5.5, and more thinking made it worse
On the new coding-agent benchmark Opus 5.5 leads at 45.5% and Astra follows at 42.8%; on one bug Astra passed 5/5 at low reasoning but only 1/5 at maximum, where it deleted a public API.
Holy. Cognition has released FrontierCode 2.0, a tougher benchmark for AI coding agents. The results are brutal: Claude Opus 5.5 leads with just 45.5%, followed by GPT-6 Astra at 42.8%. Even crazier, MORE thinking made Astra worse. On one bug, it passed 5/5 attempts at low reasoning but just 1/5 at maximum, where it deleted a public API. Claude Opus 5.5 fixed another bug but failed ALL 25 attempts across five reasoning settings because it missed a critical cleanup requirement. Even when AI fixes the bug, it can break the code around it.
This post quotes another post: x.com/cognition/status/2115132488775524793