Artificial Analysis studies how tool use affects model benchmark performance

On November 12, 2026, the evaluation organization Artificial Analysis published the results of its tool-use evaluation on VIALS-AA, its own implementation of the VIALS benchmark from Handshake AI for visual reasoning in the life sciences. The release lets readers compare models on three dimensions: task score, cost per task, and tool usage.

The results show that tool use sharply lifts performance across models, nearly doubling average scores on VIALS-AA from roughly 25% to roughly 47%. When tools are removed, Claude Fable 5.2 scores below Claude Haiku 5.5 even though the latter is evaluated with tools. The benefit is largest on counting tasks, which call for an exact tally of many small objects. With tools, models can crop images, read pixel-level intensities, and manipulate the images further to separate objects and count them in code; without tools, they must estimate from what they see.

Most of the gains from tools go to models that score lowest without them. The ten models scoring below 31% without tools average about 42 tool calls per task, while the seven higher-scoring models average about 23. MiMo-V2.6-Pro records the biggest gain of any model, scoring just 5% without tools. Enabling tools also raises the cost per task for every model, but by very different amounts: Fable rises from $0.73 to $4.20 per task, a 5.8x difference; Astra moves from $0.21 to $0.27, and Sol from $0.05 to $0.06. The material states that the largest multiples are shown by Qwen3.8 Max (0902, xhigh). So far Artificial Analysis has disclosed the with-tools results, and a fuller comparison of models without tools remains pending.