Model & task profile
—input tokens / task
—screenshot share of bill
—cost / task
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.
How this calculator works
The Computer Use Agent Cost Calculator estimates what it costs to run a computer-use or browser-automation agent, where each step feeds a fresh screenshot into the model as vision tokens. You enter your model, the number of tasks per month, the input and output prices per million tokens, how many screenshots you keep in context, and the percentage of failed or retried tasks. From these, the tool projects your monthly bill. The key driver it captures is context accumulation: because an agent loops through many steps and each screenshot is expensive, retained images stack up across a task, so cost grows with steps, image retention, and retries rather than with task count alone.
The practical lever this calculator highlights is screenshot retention. Keeping every prior screenshot in context makes each step re-pay for all earlier images, so a keep-last-N strategy — retaining only the most recent screenshots — can cut bills roughly two to three times without much loss of useful state. The trade-off to watch is that trimming too aggressively can strip context the agent needs, raising the failed/retried rate, and retries are billed too. Adjust the retention and retry inputs together to find the point where you spend the least while keeping tasks reliable.
Frequently asked questions
Why are computer-use agents so much more expensive than chat completions?
A computer-use agent works in a loop: screenshot the screen, send it to the model as vision input tokens, receive a click or keystroke action, execute it, screenshot again. Each screenshot costs roughly 1,000-1,600 input tokens depending on resolution, and a single task routinely takes 10-30 steps. If the agent keeps prior screenshots in context, input tokens grow with every step — a 15-step task keeping full history can burn well over 150K input tokens, hundreds of times a typical chat completion. The vision tokens, not the model's text output, dominate the bill.
What is the keep-last-N screenshot strategy and how much does it save?
Instead of resending every previous screenshot on each step, the agent keeps only the most recent N screenshots in context (Anthropic's reference implementation defaults to trimming old ones). For a 15-step task at 1,200 tokens per screenshot, keeping full history sends about 144K screenshot tokens across the task, while keeping the last 3 sends about 50K — roughly a 2.9x reduction in the dominant cost component, usually with no loss in task success because old screenshots of stale screen states rarely help the next action. It is the single highest-leverage cost lever in computer use.
How do Claude computer use and OpenAI computer-use-preview pricing compare?
Both bill standard token rates on the loop: Claude Sonnet-class computer use at $3/M input and $15/M output (with a promotional $2/$10 rate in effect through August 2026), OpenAI computer-use-preview at $3/M input and $12/M output. Since input vision tokens dominate the workload, the input rate matters most and the two are close — the real cost difference comes from steps per task and screenshot retention strategy, not the provider. Claude's computer-use tool also adds roughly 700-1,200 fixed system/tool tokens per request, which this calculator models as per-step overhead.