Claude computer use: the small settings that decide accuracy

Lucas Gonzalez and Luca Weihs of Anthropic wrote a technical guide to computer and browser use with Claude, and it is specific in a way most such posts are not: it is built around numbers, not principles. A few you can apply today: start at 1280x720, or 1080p for Opus 4.7, and always downscale images before sending them, because the API quietly degrades anything past its limits (1,568 px on the long edge for Claude 4.6, 2,576 px for Opus 4.7).

One detail is easy to miss. Putting the text instruction before the image in the messages array improves accuracy. On thinking effort, the 4.6 models hit near-peak performance at "medium" for half the token cost of "high", while Opus 4.7 does best at "high" and gains little at "max". Even low effort beats no thinking, because fewer mistakes mean fewer retries.

The caching section has a trap worth knowing. A rolling screenshot buffer on its own breaks the cache. The fix is to keep the three most recent screenshots but prune in batches of 25, and to set four cache breakpoints: one on the stable prefix and three on the latest tool results. For dense interfaces, turn on enable_zoom and use the keyboard for tiny targets like checkboxes. The full list is on Anthropic's blog.

Why it matters

If you are building computer-use agents, these settings decide whether your agent is accurate or expensive. The caching detail alone is the difference between a cheap loop and one that silently re-pays for context on every single step.