Tools matter most for counting tasks, which need an exact tally of many small objects. With tools, models can crop, pull pixel-level intensities and further manipulate images to separate objects and count them in code. Without them, they have to estimate from what they see.
On the 23 cell counting and quantification tasks in VIALS-AA, Fable drops from a 65% pass^3 rate with tools to just 9% without. Claude Opus 5.5 (max) falls from 57% to 4%, and Claude Sonnet 5.5 (max) falls from 65% to 13%.
GPT-6 Astra (max) and GPT-6.1 Sol (max) hold up better: Astra passes 61% with tools and 35% without, and Sol falls from 61% to 39%. Some domains depend less on tools: on the 23 phylogenetics tasks, Astra passes 78% with tools and 83% without.