Cognition announced a new coding benchmark called FrontierCode 2.0 that evaluates a model’s ability to make judgments on underspecified prompts.
In a post on X, Cognition said FrontierCode 2.0 expands on the earlier version by including more diverse tasks, larger scopes and higher mergeability standards. The benchmark deliberately leaves engineering decisions ambiguous so models must demonstrate judgment beyond simply meeting a specification.
Cognition added that the benchmark is “far from saturated,” noting that the top‑performing model achieved a score of just 45 percent.
DevinAI, the official account for Cognition’s Devin platform, echoed the announcement, describing FrontierCode 2.0 as “our new coding benchmark that measures a model’s ability to go beyond the spec” and repeating the 45‑percent top score.
Sources
Fan coverage from Devin Central — not an official Cognition announcement. Devin is a trademark of Cognition.