Guide · data from our published rounds · updated 2026-11-21
Browser automation tools, compared with data we measured
Most "best browser automation tool" lists are opinions. This page only uses numbers from benchmarks we ran and published with their scripts, plus licence and activity facts from each project's GitHub repository. Tools we haven't run are marked as such.
Scripted libraries: Playwright vs Puppeteer vs Selenium
Same 10 tasks, 10 runs each, all three driving one Chromium binary (method and per-task results).
| Library | Runs passed | Sum of task medians | Median launch | Licence |
|---|---|---|---|---|
| Playwright | 100/100 | 18,915 ms (+1.7%) | 26 ms | Apache-2.0 |
| Puppeteer | 100/100 | 18,600 ms (fastest) | 46 ms | Apache-2.0 |
| Selenium (WebDriver BiDi) | 100/100 | 19,195 ms (+3.2%) | 116 ms | Apache-2.0 |
What the data says: all three completed every run, and the totals are within a few percent of each other on these pages. Speed is not a strong reason to choose between them here; browser support, language bindings and API style matter more. See Playwright vs Puppeteer.
Browser engines: Chromium vs Firefox vs WebKit (Playwright)
Same tasks, Playwright driving its bundled engines in the official Docker image (details).
| Engine | Runs passed | Sum of task medians | Median launch |
|---|---|---|---|
| Chromium | 100/100 | 19,687 ms (fastest) | 17 ms |
| Firefox | 100/100 | 21,396 ms (+8.7%) | 170 ms |
| WebKit | 100/100 | 22,858 ms (+16.1%) | 24 ms |
AI browser agents
So far we have run one agent setup: Claude Code in headless mode with Playwright MCP, on 10 tasks with checkable answers (full results and disclosure). Browser Use, Stagehand and Skyvern are next.
| Setup | Correct | Median time per task |
|---|---|---|
| Scripted Playwright (no model) | 50/50 | 1.3 s |
| Claude Code + Haiku | 50/50 | 11.1 s |
| Claude Code + Sonnet | 50/50 | 12.7 s |
| Claude Code + Opus | 30/30 | 12.9 s |
What the data says: on simple, well-defined tasks the agent got every answer right but took roughly ten times longer per task than a script someone had already written. An agent saves writing the script; it doesn't save run time. For the design differences between agent frameworks, see Browser Use vs Stagehand.
Every tool we track
Browser automation library
- Playwright — Apache-2.0, 100,730 GitHub stars · benchmarked
- Puppeteer — Apache-2.0, 95,879 GitHub stars · benchmarked
- Selenium — Apache-2.0, 34,603 GitHub stars · benchmarked
AI browser agent framework
- Browser Use — MIT, 121,661 GitHub stars · not yet benchmarked
- Stagehand — MIT, 26,110 GitHub stars · not yet benchmarked
- Skyvern — AGPL-3.0, 23,485 GitHub stars · not yet benchmarked
Hosted / self-hostable browser infrastructure
- Steel Browser — Apache-2.0, 7,999 GitHub stars · not yet benchmarked
- Browserless — see repository licence, 13,956 GitHub stars · not yet benchmarked
Crawling & scraping toolkit
- Crawlee — Apache-2.0, 26,511 GitHub stars · not yet benchmarked
- Firecrawl — AGPL-3.0, 205,977 GitHub stars · not yet benchmarked
Headless browser engine
- Lightpanda — AGPL-3.0, 42,299 GitHub stars · not yet benchmarked
GitHub data fetched 2026-11-21. All scripts and raw results: github.com/RichardDu1/browsebench-harness.