Benchmark results

June 24, 2026 · View on GitHub

Run run_20260622_205455_c5fc. 19 login-free, real-web tasks, each run 3 times with the extension off (baseline) and on (guarded). Agent: gpt-5-mini via the Browserbase Model Gateway. Judge: claude-sonnet-4-6. Reproduce with the harness.

Headline

MetricBaseline (off)Shielded (on)
Total tokens2.33M2.07M (−11%)
Task success81% (46/57)91% (52/57)

Per-task

Sorted by token change. Negative = fewer tokens with the shield on (good), positive = more. "Runs passed" is out of 3, off → on.

TaskRuns passed (off → on)Tokens offTokens onTokens vs baseline
github-trending1/3 → 3/343,97912,702−71%
weather-nyc3/3 → 3/3121,05745,587−62%
weather-nyc-week-coldest3/3 → 2/3112,67847,747−58%
target-air-fryer1/3 → 3/386,02442,270−51%
github-react-bug-newest3/3 → 3/378,76045,640−42%
etsy-wallet-shop-sales2/3 → 3/3322,010204,605−37%
mdn-array-map3/3 → 3/356,78641,883−26%
python-docs-len3/3 → 3/3120,70789,846−26%
amazon-headphones3/3 → 3/3150,996122,002−19%
google-shopping-keyboard2/3 → 3/327,65523,014−17%
linkedin-dev-jobs2/3 → 3/326,16723,145−12%
mdn-includes-vs-flat-firefox1/3 → 2/3510,153473,456−7%
arxiv-recent-cs-ai3/3 → 3/328,15527,924−1%
wikipedia-einstein-advisor3/3 → 3/3246,699256,646+4%
craigslist-nyc-bikes2/3 → 3/361,39168,893+12%
hn-top-title3/3 → 3/315,39217,494+14%
ikea-billy-cheapest-white2/3 → 2/3158,220208,214+32%
wiki-claude3/3 → 3/3104,139173,514+67%
company-about-stripe3/3 → 1/360,305147,626+145%

Takeaways

  • The aggregate is the trustworthy claim. As you can see from the table it varies wildly across sites and across runs, but 11% savings is a safe bet. As with everything, your mileage may vary so we encourage you to run the harness on scraping taks you're doing today.
  • Rows where either side didn't pass all 3 runs are a softer comparison — a failed run can burn very different token counts, so the percentage isn't strictly like-for-like. The cleanest wins are the 3/3 → 3/3 rows (weather, GitHub issues, MDN, Python docs, Amazon).
  • We leave the regressions in on purpose. company-about-stripe got worse and used 145% more tokens; we kept it rather than quietly dropping it. A benchmark that only shows wins isn't a benchmark.
  • This is gpt-5-mini only, on the non-blocked slice of sites. A directional signal, not a published paper.