Browser agent cost per task: Claude Sonnet 5.5 on 23 real sites
Across 23 real-site tasks, a Claude Sonnet 5.5 browser agent cost a median of $0.15 per run. Costs ranged from $0.01 to $0.59. A Wield recipe run has no model in the loop, so its model cost is $0. Figures are as of 7 November 2026, from a benchmark that is still running, and they are list-price model costs only.
Key finding
- Median cost per run: $0.15 for the browser agent (0.1476 before rounding). The mean is $0.19.
- Range: $0.01 (zillow.com) to $0.59 (bahn.de, one run).
- Wield: $0 per run, because no model runs in the loop.
Cost per run by task
| Site | Task | Agent cost per run | Agent runs passing | Wield cost |
|---|---|---|---|---|
| bahn.de | search-connections | $0.59 | 0 of 1 | $0 |
| autoscout24.de | search-cars | $0.54 | 5 of 5 | $0 |
| willhaben.at | search-marketplace | $0.47 | 4 of 4 | $0 |
| doctolib.de | search-practitioners | $0.38 | 2 of 2 | $0 |
| check24.de | compare-electricity | $0.26 | 5 of 5 | $0 |
| flixbus.de | search-trips | $0.23 | 3 of 3 | $0 |
| kleinanzeigen.de | search-listings | $0.21 | 2 of 2 | $0 |
| vinted.de | search-items | $0.20 | 7 of 8 | $0 |
| airbnb.com | search-stays | $0.19 | 4 of 4 | $0 |
| doctolib.de | get-slots | $0.16 | 1 of 2 | $0 |
| amazon.com | search-products | $0.16 | 4 of 4 | $0 |
| jameda.de | search-doctors | $0.15 | 9 of 9 | $0 |
| jameda.de | get-slots | $0.13 | 9 of 9 | $0 |
| zalando.de | search-products | $0.12 | 0 of 9 | $0 |
| walmart.com | search-products | $0.12 | 0 of 9 | $0 |
| chefkoch.de | search-recipes | $0.10 | 8 of 8 | $0 |
| opentable.com | search-restaurants | $0.08 | 0 of 9 | $0 |
| eventim.de | search-events | $0.06 | 0 of 9 | $0 |
| lieferando.de | search-restaurants | $0.05 | 0 of 9 | $0 |
| idealo.de | price-lookup | $0.03 | 0 of 9 | $0 |
| immobilienscout24.de | search-rentals | $0.02 | 0 of 9 | $0 |
| mobile.de | search-cars | $0.02 | 0 of 9 | $0 |
| zillow.com | search-homes | $0.01 | 0 of 9 | $0 |
Cost when a run passes vs when it is blocked
The 13 tasks where the agent passed at least once have a median cost of $0.20 and a mean of $0.24. The 10 tasks where every run was blocked have a median of $0.06 and a mean of $0.11. Blocked runs end early, so they cost less.
The cheapest rows are mostly failed runs. Do not read $0.01 as an efficient result.
Cost per successful result
The table does not give cost per passing run, because cost is averaged over passing and failed runs. Two passing tasks show the spread. autoscout24.de cost $0.54 per run and passed 5 of 5. willhaben.at cost $0.47 per run and passed 4 of 4. zalando.de cost $0.12 per run and passed 0 of 9.
Per run, a passing agent task cost roughly $0.20 to $0.50 in these results. A Wield recipe run cost $0 in model tokens for the same tasks.
Method and caveats
- Agent: Browser Use 0.13.14 with Claude Sonnet 5.5, stock settings, in a fresh headless Chromium with an empty profile. The median run used 7 steps and 42,788 input tokens.
- List price only. Costs use the model's list price as logged. They exclude browser hosting. Wield's own cost is $0 in model tokens, but it uses the user's own browser.
- Small samples. bahn.de's $0.59 comes from a single run. The benchmark is paced at about 30 site requests per hour.
- Not a bill. Do not extrapolate these numbers to monthly costs.
FAQ
How much does a browser agent cost per task?
In our benchmark of 23 tasks, the median was $0.15 per run. Costs ranged from $0.01 to $0.59. These are list-price model costs for one browser agent setup.
Why does a blocked run still cost money?
The agent spends model tokens before it stops. In our data, a blocked run still cost between $0.01 and $0.59, depending on how many steps it took before the block.
What does a Wield run cost?
A Wield recipe run has no model in the loop, so its model cost is $0. Public recipe runs and unlimited local runs are free. Paid plans cover the recipe builder, private recipes and team features, and are listed on pricing.