How fast can an AI agent use a website?
An AI agent that uses a website through a recipe gets a result in a median of 0.9 s. An agent that drives a cloud browser with screenshots and clicks needs a median of 78 s for the same kind of task. In our benchmark of 24 tasks on 21 sites, as of 7 Nov 2026, the recipe was about 85 times faster at the median. The browser agent also cost $0.18 per run in model tokens. The recipe cost $0.
Median times
These are the pooled numbers from the benchmark, with equal weight per task:
| Wield recipe | Browser agent | |
|---|---|---|
| Median time per run | 0.9 s | 78 s |
| Success across tasks | 85% | 52% |
| Model cost per run | $0 | $0.18 |
| Pooled runs passing | 183 of 208 | 63 of 147 |
The browser agent was Browser Use 0.13.14 driving Claude Sonnet 5.5 with stock settings in a fresh headless Chromium. The recipe ran through the Wield extension in an everyday browser. Both sides had to pass the same machine check.
Per-task numbers
The table shows 8 of the 24 tasks. Times are p50 from the benchmark.
| Site | Task | Wield p50 | Browser agent p50 | Browser-agent runs | Browser agent blocked by bot protection |
|---|---|---|---|---|---|
| bahn.de | Train connections with prices | 2,127 ms | 675,891 ms | 1 | Yes |
| flixbus.de | Bus trips with prices | 791 ms | 41,150 ms | 3 | No |
| kleinanzeigen.de | Classified listings search | 759 ms | 125,287 ms | 2 | No |
| check24.de | Electricity tariff comparison | 1,544 ms | 103,392 ms | 5 | No |
| lieferando.de | Restaurant search | 1,796 ms | 54,783 ms | 9 | Yes |
| jameda.de | Free appointment slots | 211 ms | 46,005 ms | 9 | No |
| doctolib.de | Free appointment slots | 345 ms | 70,636 ms | 2 | No |
| chefkoch.de | Recipe search | 414 ms | 39,527 ms | 8 | No |
Wield passed 9 of 9 runs on most of these tasks. The doctolib.de slot task passed 8 of 8. "Blocked" means the cloud browser hit a bot check in our runs. It does not mean the site always blocks agents.
Why recipes are fast
A recipe makes the same HTTP requests the site's own page makes. It returns structured data. No model reads a page between steps, and no screenshot is taken, so there is no waiting on a model. A browser agent does the opposite: it reads the page, decides, clicks, and reads again. Each step adds model time and tokens.
The trade-off is coverage. A recipe exists only for a task someone has written, and it breaks when the site changes its requests. Wield checks public read recipes daily and marks failing ones as degraded. A browser agent can attempt any page, but it is slower and costs more per task.
Caveats
- These are benchmark medians, not a promise for every task. The benchmark is preliminary and still running.
- Some numbers in our live checks come from single verification runs in our recipe status notes, not from the benchmark. For example, Chefkoch returned in 0.2 s and willhaben in 0.45 s in single checks on 7 Nov 2026. Do not treat them as medians.
- The bahn.de recipe has a pending live check in our recipe status notes. Read its row with that in mind.
- Our own misses are part of the data. Airbnb passed 8 of 9 runs. mobile.de passed 5 of 9.
- Use "about 85 times faster at the median" as the only speed ratio. Do not extend it to other tasks.
FAQ
What is the speed of a browser agent?
In our benchmark, the median browser-agent run took 78 s across 24 tasks. Single tasks ranged from about 15 s to about 11 minutes, and the slowest runs were ones that bot protection blocked. Full method is on the benchmark page.
Does a recipe run cost anything?
No model tokens are used when a recipe runs, so the model cost per run is $0. The run happens on your computer, through your browser.
Can I see the raw data?
Per-task rows are in the benchmark. The Deutsche Bahn page shows one task in detail: train connections benchmark.