Model & task profile
Screenshot retention — the biggest cost lever
Same task, same steps — only how many previous screenshots stay in context changes. Old screenshots of stale screen states rarely improve the next action, but they always bill as fresh input tokens.
| Strategy | Input tok / task | Cost / task | Monthly | vs keep-all |
|---|
The screenshot loop is where browser automation budgets die
A computer-use agent does not read the DOM — it looks at the screen. Every step of every task is a full round trip: capture a screenshot, send it to the model as vision input tokens, get back a click, type, or scroll action, execute it, and capture the next screenshot. At roughly 1,000-1,600 tokens per screenshot (resolution-dependent) and 10-30 steps for a routine task, the vision input alone dwarfs everything else in the request. The model's actual text output — a coordinate and a short justification — is trivially small by comparison. This inversion catches teams off guard: in chat workloads output tokens are the expensive part; in computer use, input is typically 80-95% of the bill.
Context accumulation is the multiplier on top. If the agent keeps every previous screenshot in the conversation, step 15 resends fourteen stale screenshots plus the current one, and total screenshot tokens grow quadratically with task length — a 15-step task at 1,200 tokens per screenshot burns about 144K screenshot tokens with full history versus about 50K keeping only the last three. That single retention setting changes the dominant cost component by nearly 3x, which is why Anthropic's reference implementation trims old screenshots by default. The same accumulating-context math for text-only agents is covered in our AI agent step cost calculator; screenshots just raise the stakes by an order of magnitude.
Two practical notes. First, budget for retries: computer-use agents fail and re-run tasks at meaningful rates (UI changes, timeouts, mis-clicks), and a 10-20% retry allowance belongs in any honest estimate. Second, resolution is a real lever — downscaling captures to the model's effective resolution before sending avoids paying for pixels the model resamples away anyway. If your workload is scraping-shaped rather than interaction-shaped, compare against a conventional web scraping API: when no clicking or typing is required, screenshots are an expensive way to read a page.