← All benchmarks

FrontierCS Algorithmic v1.1

Code · 2026-10-18

179 open-ended algorithmic problems, many NP-hard variants of competitive-programming tasks, whose checkers accept any valid answer and score better answers higher. Version 1.1 sets each problem's reference at 60 points and adds the natural log of further improvement, capped at 100, and rescores agent runs made in each lab's official harness. The score is the equal-weight mean over problems, with a missing submission scored 0; version 1.0 scores are not comparable.

Top models (higher is better)