Claude Opus 5.5 leads Cognition’s FrontierCode 2.0 with 45.5%

Top News

FrontierCode 2.0 evaluates how coding agents handle underspecified work in existing codebases. Cognition reports Claude Opus 5.5 at 45.5% versus GPT-6 Astra at 42.8%.

Here is what makes this useful:

• Tasks emphasize inferred requirements and thoroughness, such as fixing related instances of a bug and matching existing support for other input types.
• Blocking criteria are decisive: a solution gets zero if it fails them, even though scores otherwise aggregate rubric items.
• Tasks span 55% more repositories than FrontierCode 1.1, with reference solutions changing 67% more files on average.
• More than 20 open-source developers built tasks in repositories they maintain, spending over 40 hours per task and grading for whether changes are mergeable.

One catch: Higher reasoning effort does not reliably improve results. Extra code or file changes can hurt scores through scope violations and overthinking.

Sources: cognition.com