Benchmark: search recipes on Chefkoch with an AI agent
On Chefkoch, Wield's recipe passed 9 of 9 recipe searches with a median of 0.4 seconds and no model cost per run. A browser agent running Claude Sonnet 5.5 passed 8 of 8 runs, with a median of 40 seconds and $0.10 per run. Both sides passed every run. The difference in this task is speed and model cost.
Chefkoch Search cooking recipes with rating, difficulty and preparation time
Read-only, no loginPasses the daily check
Parameters (4)
| Metric | Wield recipe | Browser Use + Claude Sonnet 5.5 |
|---|---|---|
| Success | 100% | 100% |
| Median time | 414 ms | 40 s |
| Cost per run | $0 | $0.096 |
9 runs each, 7 November 2026. How we measure
The task
Both sides ran the same three searches, with up to 10 results each.
| Query | Search text |
|---|---|
| q1 | lasagne |
| q2 | kürbissuppe |
| q3 | vegan schnell |
Pass check: at least five recipes, each with a title and a rating.
Results (7 Nov 2026)
| Wield recipe | Browser agent | |
|---|---|---|
| Runs passing the check | 9 of 9 | 8 of 8 |
| Median time per run | 414 ms | 39.5 s |
| Model cost per run | $0 | $0.10 |
| Requests per run (median) | 1 | 187 (10 to the site) |
| Blocked by the site | 0 runs | 0 of 8 runs |
The browser agent took a median of 5 steps, 30,299 input tokens and 1,616 output tokens per run. Wield had no failed runs. Its fastest run took 181 ms and its slowest 4,916 ms.
The browser agent's log records 401 responses on one account endpoint, /benutzer/me, nine times across its runs. Those runs were not counted as blocked, and all 8 passed. The benchmark does not treat this as a block.
Why the results differ
Chefkoch offers a public JSON search gateway that allows requests from www.chefkoch.de. The Wield recipe calls it from a chefkoch.de tab in your browser, at human pace. It makes one request per search at the median and drops brand-campaign recipes by default.
The browser agent opens the search page, reads the results and moves between pages. Each of its median 5 steps waits for a model call.
Caveats
- Small and preliminary. The benchmark was still running on 7 Nov 2026. 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.
Chefkoch's terms, in plain words
Chefkoch's robots.txt disallows AI training crawlers (ClaudeBot, anthropic-ai, Claude-Web). Chefkoch's terms restrict automated use of its content. Wield runs the recipe only when you ask, one search at a time from your own browser, with one request per search. The scope is set: you start each search, one search per call, from your own browser, with no bulk collection and no purchases. Recipe texts stay on chefkoch.de through the link. Check the site's terms yourself. Wield's recipes note known restrictions.
FAQ
Does Chefkoch block browser agents?
Not in these runs. The browser agent was blocked by the site in 0 of 8 runs. Its log shows 401 responses on one account endpoint, which the benchmark did not count as a block. Other searches and times may differ, and bot protection can change at any time.
How fast is the Wield recipe for search recipes on Chefkoch?
The median run took 414 ms over 9 runs. The fastest took 181 ms and the slowest 4,916 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.