60% of tests made private
Artificial Analysis, which compares AI models under the same conditions, released version 5 of its overall "Intelligence Index" on Oct. 27 (Oct. 28 Japan time). For the first time, private tests, whose questions and answers are not made public, account for 60% of the index. Once questions circulate, models can memorize the answers, making it impossible to measure their true ability.
It newly added Cognition's "FrontierCode 2.0" (132 tasks that grade whether real code changes are of mergeable quality) and "Terminal-Bench-Science," 70 tasks that advance scientific research. The index combines four areas, knowledge work, programming, science and general, at 25% each. It also launched a "Progress Index" that compares gains in model performance over long periods.
Gap with open models grows to 12 points
The highest-scoring model whose developer plans to release its weights is Kimi K3.1 Pro from China's Moonshot AI at 46 points, and its gap with the leader widened from 3 points in the previous version to 12. However, the weights of Kimi K3.1 Pro have not yet been released. Among models whose weights are already out, the highest are DeepSeek V4.1 Flash 1017 and GLM-5.3 at 32 points, 26 points behind the leader.
Opus 5.5 has the highest cost per task
The cost per test task was $10.04 for Claude Opus 5.5, $6.24 for GPT-6 Astra and $1.27 for GPT-6.1 Sol. Opus 5.5's score varies widely with reasoning depth, from 35 points at the shallowest setting to 58 at the deepest. The median number of tokens output per task rose from 35,000 in the previous version to 59,000.
