What does a browser agent cost per task?
A browser agent pays model tokens for every step it takes, including the steps that fail. In our benchmark, Browser Use with Claude Sonnet 5.5 cost $0.184 per run on average, with a median time of 78 s. A Wield recipe cost $0 in model tokens, because no model runs during a recipe, and its median time was 0.9 s. These results are preliminary, as of 7 November 2026.
The numbers
| Measure | Browser agent | Wield recipe |
|---|---|---|
| Median time per run | 78 s | 0.9 s |
| Model cost per run (average) | $0.184 | $0 |
| Success | 52% | 85% |
| Pooled runs passing | 63 of 147 | 183 of 208 |
Cost is the model's list price per run, averaged with equal weight per task. Time is the median wall time per run. Each task has 3 fixed queries, and each query runs 3 times per side.
The cost figures count model usage only. They do not include your time, your AI app's subscription or any retries you run by hand. The time figures are wall-clock time for each run. Success is decided by a machine check, not by what the agent reports. Both sides must pass the same check for each task, such as at least one train connection on the requested date with a departure, arrival and price.
Where the cost goes
A browser agent sends the state of the page to the model at each step, and the model decides the next action. Every step adds input and output tokens. Failed steps and retries add more steps, so a blocked or looping run costs more than a clean one. A recipe skips this loop. It makes the requests the site's own page makes and returns the data.
The most expensive tasks
| Site and task | Browser agent cost per run | Browser agent success | Recipe median time |
|---|---|---|---|
| bahn.de, train connections | $0.5888 | 0% (blocked) | 2.1 s |
| AutoScout24, car search | $0.5389 | 100% | 0.9 s |
| Willhaben, marketplace search | $0.472 | 100% | 0.6 s |
The bahn.de row is the extreme case. The browser agent was blocked, produced no result and still cost $0.5888. Failed runs cost money, and that is part of the figure.
What Wield costs
Running a recipe costs no model tokens, because it runs on your computer. The extension, wield-mcp and the public recipes are free. Plans add private recipes, the recipe builder and team features. See pricing for the details.
Your AI app still uses its own model to read your request and pick the tool. Wield does not add a model charge on top of that.
FAQ
Is $0 per run realistic?
In the benchmark, yes. The recipe runs used no model, across the 24 tasks. Your AI app's own model use is separate, and it is not part of the recipe cost.
Why does the browser agent cost differ by site?
Each site asks for different steps, and some runs fail. Failed runs still cost model tokens. The bahn.de run was blocked and cost $0.5888 with no result, so it is the highest figure in the table.
Are these numbers final?
No. As of 7 November 2026, the benchmark is preliminary and still running. The figures will change as more runs finish.
Where can I see the raw data?
The benchmark page has the method and the per-task results. The raw logs hold destinations, timing, token counts and check results. They do not hold page contents.