A surprising amount of information reaches you as an image: a chart in a PDF with no underlying data, a scanned invoice, a screenshot of a dashboard someone will not give you access to. Claude Opus 5's high-resolution vision handles all three, and it maps coordinates one-to-one with actual pixels, which matters for anything involving UI.
This page is written for anyone whose data arrives as a picture. If you want the model overview first, start with the complete Claude Opus 5 guide.
| Best effort level | high |
| Context available | 1M tokens — enough for real working material, not just snippets |
| Cost | $5 per million input tokens, $25 per million output |
Why Claude Opus 5 suits visual analysis
- High-resolution input up to 2,576 px on the long edge, so dense charts and small text stay legible.
- Strong on chart, diagram, and document understanding — including reading values off a plot with no data table.
- Good at UI and frontend visual replication, which makes screenshot-to-code workflows practical.
- Coordinates map directly to image pixels, so no scale-factor arithmetic in computer-use workflows.
A workflow that works
- Send the highest resolution you have for dense material — downscaling is what causes misread axis labels.
- Give it tools to crop and re-examine. Letting it zoom into a region is more effective, and cheaper, than raising its reasoning effort.
- Ask it to transcribe before interpreting — get the numbers out first, then reason about them.
- For screenshots-to-code, describe the intent, not just the visual: what the component does, not only what it looks like.
- Downscale deliberately for cost when fidelity does not matter.
Settings to start from
Set the effort level to high. For vision, tools beat effort. Giving the model the ability to crop and verify its own reading improves accuracy more cost-effectively than turning the reasoning dial up.
Leave adaptive thinking on — it is the default on Claude Opus 5, and turning it off introduces failure modes that are not worth the saving. See effort levels explained if you want to tune this properly.
What to watch out for
- Images are not free. A full-resolution image can consume ~4,784 tokens per image. At scale, that adds up — downsample when the detail is not load-bearing.
- Verify numbers read off charts against a source where one exists.
- Old prompt workarounds may hurt. Instructions written to compensate for a weaker model's vision are worth re-testing; several are now counterproductive.
The verdict
One of the most practically useful upgrades in the model, and the one people most often forget they have.
If cost is the constraint, read how to cut Claude Opus 5 costs before downgrading — routing and caching usually beat switching models.
More on Claude Opus 5
Start with the complete Claude Opus 5 guide for the overview, or go deeper:
Ready to go deeper?
Read the full Claude Opus 5 guideFrequently Asked Questions
What image resolution does Claude Opus 5 support?
Up to 2,576 pixels on the long edge, which is the high-resolution tier. That is enough for dense charts, small print in scanned documents, and full-page screenshots to remain legible without pre-processing.
How much do images cost in tokens?
A full-resolution image can use up to roughly 4,784 tokens. At $5 per million input tokens that is a fraction of a cent per image, but it compounds in high-volume pipelines — downsample when the extra detail is not needed.
Can it read a chart with no underlying data?
Yes — reading values off a plotted chart is one of its stronger visual skills. The reliable technique is to ask it to transcribe the values first and interpret second, and to give it a crop tool so it can zoom into a dense region rather than guessing from the full image.
