Benchmark: search restaurants on OpenTable with an AI agent
Wield's recipe found restaurants on OpenTable in 9 of 9 runs, with a median of 1.9 s and no model cost. A Claude browser agent passed 0 of 9 runs, with a median of 27.6 s and $0.08 per run. The benchmark flagged all 9 agent runs as blocked by the site. These results are preliminary, as of 7 November 2026.
OpenTable Find restaurants to book on OpenTable for a date, time and party size
Read-only, no loginRuns in your browser only
Parameters (5)
| Metric | Wield recipe | Browser Use + Claude Sonnet 5.5 |
|---|---|---|
| Success | 100% | 0% |
| Median time | 1.9 s | 28 s |
| Cost per run | $0 | $0.077 |
9 runs each, 7 November 2026. How we measure
The task
Each side ran the same three searches. A run passes when it returns at least five restaurants with a name and a cuisine.
| Query | Parameters |
|---|---|
| q1 | query = sushi, city = New York, date = 2026-11-20, time = 19:00, party_size = 2 |
| q2 | query = italian, city = San Francisco, date = 2026-11-21, time = 20:00, party_size = 2 |
| q3 | query = steakhouse, city = Chicago, date = 2026-11-22, time = 19:30, party_size = 4 |
Results
| Wield recipe | Browser agent | |
|---|---|---|
| Runs passing the check | 9 of 9 | 0 of 9 |
| Median time per run | 1.9 s (1,940 ms) | 27.6 s (27,616 ms) |
| Model cost per run | $0 | $0.08 |
| Requests per run (median) | 1 | 17 (to the site: 0) |
| Blocked by the site | 0 runs | 9 of 9 runs |
The browser agent took a median of 9 steps per run, with 23,686 input tokens and 1,489 output tokens. Wield had no failed runs. Its fastest run took 1.5 s and its slowest 11.0 s.
Why the results differ
Wield sends one search request from a tab in the user's own browser, at human pace. No model decides the steps, so the run costs no tokens.
The browser agent ran in a fresh headless Chromium with an empty profile. Its requests failed with navigation errors, and the log holds no HTTP 401, 403 or 429 status. Our recipe notes say OpenTable drops connections that do not come from a normal browser. We did not confirm the cause of each failure.
Caveats
- Small sample. Each side has at most 9 runs, and the benchmark is still running. It is paced at about 30 site requests per hour.
- Blocked means the benchmark's flag. It covers an HTTP 401, 403 or 429 from the site, or the agent reporting a block. Here the agent reported site unavailable.
- Wield does not solve bot protection. If OpenTable shows a check, Wield stops with
site_challenge, and the user solves it in their own tab.
Terms of OpenTable
OpenTable's terms prohibit automated access. The recipe makes one search at a time that the user starts, from the user's own browser. Reservations are made by the user on OpenTable. Check the site's terms before you use any automation.
FAQ
Does OpenTable block browser agents?
In our runs, yes. A Claude browser agent in a fresh headless Chromium was flagged as blocked in all 9 runs. Wield's recipe passed 9 of 9. The result comes from one agent setup on one date.
How fast is the Wield recipe for search restaurants on OpenTable?
In our benchmark as of 7 November 2026, the median run took 1.9 s. The fastest run took 1.5 s and the slowest 11.0 s.
Can I rerun this benchmark?
The method and per-task rows are on the benchmark page. The raw logs hold destinations, timing, tokens and check results, not page contents.