← All benchmarks

FrontierCS 2

Code · 2026-10-20

48 open-ended computer-science R&D tasks in 13 domains, each built by deleting a published paper's contribution from the code its authors released. An agent works in the repository for up to five hours with development-suite feedback, and a run passes only if its chosen final submission beats the authors' reference code on every checked metric and every hidden workload. The score is the share of an agent's 48 single runs, one per task, that pass.

Top models (higher is better)