Artificial Analysis has released version 5 of its Intelligence Index, changing the mix of tests used to compare frontier AI models. The new index gives equal 25% weight to knowledge work, coding, science and general capability, with multiple benchmarks contributing to each category.
Interconnects argues that the redesign highlights a deeper problem: modern models have jagged skill profiles that resist compression into one number. A model that performs well on repetitive coding tasks may not be equally strong at open-ended knowledge work, while the tools and execution harness around a model can materially affect results. Aggregating those differences is useful for quick comparisons, but every weighting choice also embeds a judgment about which uses matter most.
The analysis does not dismiss independent evaluation. It says these groups provide a valuable check on vendor claims and can influence how a large market understands progress. Its practical warning is narrower: users should examine category-level results and the exact testing setup rather than treating a leaderboard position as a universal measure of quality. As model capabilities broaden, separate evaluations for distinct jobs may be more informative than a single intelligence score.