OpenAI released GPT-6.1 Terra in the API on October 26, 2026, with image input for visual tasks. If you try a new model on screenshots, recheck its supported image settings and actual usage instead of copying another model’s token calculation. The comparison below is a small test you can run; no benchmark or universal saving is claimed.
An image sent for analysis consumes input tokens. File size alone does not tell you the charge: the selected model, image dimensions and supported detail setting affect processing.
Send the necessary evidence
For a question about one table, crop to that table while retaining its headings and units. Do not shrink the entire screen until the numbers become unreadable. If the task depends on a footnote, include it in the crop or send a second clearly labelled image.
Compare a full-screen image with a relevant crop using the same question and model. Record reported input usage and whether the answer correctly reads two small values. This is a proposed comparison, not a claim that a particular crop always saves money.
Verify before scaling
Check current model-specific image rules and actual usage on a small approved sample. A lower detail setting can lose information; more detail does not guarantee correct reading. Never send unrelated account details simply because they share the screenshot.