← All benchmarks

FrontierCS Algorithmic v1.0 Agents (188 problems)

Code

FrontierCS version 1.0 agent results on 188 open-ended algorithmic problems, scored under the original rule that gave many problems full marks at the reference. Agents ran through Harbor for up to five hours, could submit repeatedly for feedback and kept their best score; the score is the mean over problems. A diagnostic, because the published board also holds runs on earlier, smaller problem sets.

Top models (higher is better)