I’m testing OpenAI vision models on dashboard screenshots where the chart is readable to a person, but the axes, legend, and data labels are packed close together. They usually identify the general trend; I’m less sure about exact values and which series a label belongs to.

Has anyone evaluated this with real charts? I’d be interested in whether performance changes much between bar charts and line charts, and how you handle cases where the model sounds confident but may have misread a number. My main concern is using the results in a report without a person checking every value.