Zoom Tool Dramatically Boosts Vision Model Accuracy on Chartography Benchmark
Anthropic's Claude developers demonstrated that equipping vision models with a zoom tool significantly improves their performance on the Chartography benchmark. Specifically, Fable 5's accuracy rose from 29% to 73%, while Sonnet 5's accuracy increased from 13% to 44% when using the tool. Dense, professional charts like Sankey diagrams or Kaplan-Meier curves are notoriously difficult for AI to read due to tiny details. This technique shows that giving models tool-use capabilities like zooming can bypass physical resolution limits without needing expensive retraining or fine-tuning. The evaluation was conducted on the Chartography benchmark, which consists of 100 questions over dense real-world charts. Anthropic has shared an official cookbook containing the implementation details for developers to apply this zoom tool technique.
## BACKGROUND
Chartography is a benchmark developed by Surge AI to test how well frontier AI models can perceive, interpret, and reason from complex, domain-specific charts used by professionals in medicine, finance, and engineering. Vision-language models often struggle with these charts because high-density information gets lost when images are resized to fit the model's standard input resolution.