Benchmark: search products on Amazon with an AI agent
On Amazon, Wield's recipe passed 9 of 9 product searches with a median of 1.3 seconds and no model cost per run. A browser agent running Claude Sonnet 5.5 passed 4 of 4 runs, with a median of 45 seconds and $0.16 per run. Both sides returned at least five named, priced products, and neither side was blocked.
Amazon Search Amazon products with price, rating and Prime
Read-only, no loginRuns in your browser only
Parameters (4)
| Metric | Wield recipe | Browser Use + Claude Sonnet 5.5 |
|---|---|---|
| Success | 100% | 100% |
| Median time | 1.3 s | 45 s |
| Cost per run | $0 | $0.161 |
9 runs each, 7 November 2026. How we measure
The task
Both sides ran the same three searches.
| Query | Search text |
|---|---|
| q1 | usb c charger |
| q2 | noise cancelling headphones |
| q3 | mechanical keyboard |
Pass check: at least five products, each with a name and a price.
Results (7 Nov 2026)
| Wield recipe | Browser agent | |
|---|---|---|
| Runs passing the check | 9 of 9 | 4 of 4 |
| Median time per run | 1.3 s | 45.2 s |
| Model cost per run | $0 | $0.16 |
| Requests per run (median) | 1 | 320 (33 to the site) |
| Blocked by the site | 0 runs | 0 of 4 runs |
The browser agent took a median of 5 steps, 59,485 input tokens and 2,729 output tokens per run. Its log records no HTTP 401, 403 or 429 response from the site. Wield had no failed runs. Its fastest run took 875 ms and its slowest 5,376 ms.
Why the results differ
The Wield recipe makes one request per search at the median. It runs from an amazon.com tab in your own browser, at human pace, and drops sponsored cards. Prices follow the visitor's currency, so a run from Germany returns euros.
The browser agent opens pages, reads them and acts step by step. Each of its median 5 steps waits for a model call. The agent also sent 320 requests per run at the median, 33 of them to the site.
Caveats
- Small and preliminary. The benchmark was still running on 7 Nov 2026. The browser agent side has 4 runs. Pacing at about 30 site requests per hour keeps counts low on some sides.
- Cost is model cost only. Wield's $0 covers model cost. No model runs in the loop when a recipe runs.
- Live numbers. The live benchmark row on this page comes from
bench/results.json.
Amazon's terms, in plain words
Amazon's robots.txt disallows Claude-User (AI agents acting for users) and Anthropic's crawlers on all paths. Amazon's Conditions of Use restrict robots and data mining. Wield runs the recipe only when you ask, one search page per call, from your own browser. The scope is set: you start each search, one search per call, from your own browser, with no bulk collection and no purchases. Buying stays on amazon.com. Check the site's terms yourself. Wield's recipes note known restrictions.
FAQ
Does Amazon block browser agents?
Not in these runs. The browser agent was blocked by the site in 0 of 4 runs, and its log records no 401, 403 or 429 response. Other queries and times may differ, and bot protection can change at any time.
How fast is the Wield recipe for search products on Amazon?
The median run took 1.3 seconds (1,306 ms) over 9 runs. The fastest took 875 ms and the slowest 5,376 ms. All 9 runs passed the check.
Can I rerun this benchmark?
Yes. The method is on the benchmark page, and the query table above is the full input. The raw per-run logs keep destinations, timing, tokens and checks. They do not keep page contents.