Expanded to 322 coding AI evaluation tasks... three test scores given equal weight
Original headline코딩 AI 평가 과제 322개로 확대…세 시험 점수 같은 비중
Not investment advice. Muhabbit scores how much a story may matter, not which way to trade.
Artificial Analysis expanded its coding-agent benchmark from 290 to 322 tasks in the v2.1 methodology of its Coding Agent Index, adding tasks to the repository-editing FrontierCode-AA 2.0 suite (now 132 tasks) and revising the grader for the command-line Terminal-Bench 4.1 suite (66 tasks), alongside 124 tasks in the repository-reading SWE-Atlas-QnA-AA set. The update also changes scoring: the three test scores are now weighted equally rather than pooled into a single average, so suites with more tasks no longer carry outsized influence. The revision matters because it reshapes how coding agents are ranked on repository editing, multi-step terminal work and codebase question answering. What to watch next is how agent leaderboard standings shift under the equal-weight scheme and whether other evaluators adopt similar per-suite normalization.
Muhabbit's AI wrote this summary in its own words from TokenPost's report and translated it from Korean. It can get details wrong, so check the source.
Named in this story
Artificial Analysis, SWE-Atlas-QnA-AA, Terminal-Bench 4.1, FrontierCode-AA 2.0