Benchmark
axuse vs Playwright MCP. Every task, side by side.
Both tools drove the same AI model through the same 23 browser tasks, 71 runs each. Everything below is a median or a total across those runs. axuse used 21% less time, 23% fewer tokens and 40% fewer tool calls overall, and both passed every run.
| Across all 71 runs per tool | Playwright MCP | axuse | Difference |
|---|---|---|---|
| Tasks passed | 71/71 | 71/71 | same |
| Median time per task | 16 s | 13 s | −19% |
| Total time | 1493 s | 1183 s | −21% |
| Median tokens per task | 168k | 144k | −14% |
| Total tokens | 15.38 M | 11.84 M | −23% |
| Median tool calls per task | 8 | 5 | −38% |
| Total tool calls | 673 | 402 | −40% |
Task by task
Median per task, lower is better and shown in bold. axuse is equal or lower on time in 21 of 23 tasks and on tool calls in 23 of 23. It is not always ahead: Playwright MCP is faster on t5, t10 and uses fewer tokens on t1, t3, t8, t9, t12, t13, t14, t15, t17.
| Task | What it tests | Time | Tokens | Tool calls | Passed | ||||
|---|---|---|---|---|---|---|---|---|---|
| Playwright | axuse | Playwright | axuse | Playwright | axuse | Playwright | axuse | ||
| t1 | Click to reveal a code | 13 s | 9 s | 109k | 113k | 5 | 4 | 3/3 | 3/3 |
| t2 | Fill a form (text, email, select, checkbox) | 14 s | 10 s | 135k | 91k | 7 | 3 | 3/3 | 3/3 |
| t3 | Look up a price in a 150-row table | 9 s | 8 s | 74k | 94k | 3 | 3 | 3/3 | 3/3 |
| t4 | Four-step wizard with server-side checks | 29 s | 16 s | 353k | 198k | 16 | 7 | 3/3 | 3/3 |
| t5 | Click inside a cross-origin iframe | 12 s | 13 s | 115k | 112k | 5 | 4 | 3/3 | 3/3 |
| t6 | Settings inside nested shadow DOM | 16 s | 11 s | 158k | 118k | 7 | 4 | 3/3 | 3/3 |
| t7 | Like one post in an infinite-scroll feed | 39 s | 30 s | 290k | 170k | 12 | 6 | 5/5 | 5/5 |
| t8 | Use a hover-only menu | 13 s | 12 s | 123k | 148k | 7 | 5 | 3/3 | 3/3 |
| t9 | Answer a confirm() then a prompt() dialog | 15 s | 13 s | 157k | 178k | 7 | 6 | 3/3 | 3/3 |
| t10 | Open a link in a new tab and come back | 17 s | 23 s | 220k | 144k | 10 | 5 | 3/3 | 3/3 |
| t11 | Wait for a 5 s job, then download | 26 s | 18 s | 197k | 168k | 9 | 6 | 3/3 | 3/3 |
| t12 | Drag a card between kanban columns | 11 s | 11 s | 110k | 115k | 5 | 4 | 3/3 | 3/3 |
| t13 | Write into a rich-text editor | 14 s | 11 s | 116k | 118k | 6 | 4 | 3/3 | 3/3 |
| t14 | Pick from an autocomplete combobox | 21 s | 20 s | 205k | 225k | 11 | 8 | 3/3 | 3/3 |
| t15 | Debounced search plus pagination | 15 s | 15 s | 168k | 205k | 7 | 7 | 3/3 | 3/3 |
| t16 | Click inside a canvas (needs vision) | 19 s | 12 s | 156k | 118k | 8 | 4 | 3/3 | 3/3 |
| t17 | Test a local dev build end to end | 20 s | 18 s | 190k | 215k | 10 | 8 | 3/3 | 3/3 |
| t18 | Shop checkout with variants and a coupon | 42 s | 28 s | 641k | 390k | 26 | 12 | 3/3 | 3/3 |
| t19 | Approve one invoice in a 1,200-row grid | 34 s | 21 s | 368k | 148k | 14 | 5 | 3/3 | 3/3 |
| t20 | Choose dates in two date pickers | 38 s | 24 s | 586k | 405k | 21 | 13 | 3/3 | 3/3 |
| t21 | Sign-up form with validation errors | 20 s | 13 s | 222k | 118k | 10 | 5 | 3/3 | 3/3 |
| t22 | Fill a form nested in an embedded widget | 14 s | 12 s | 136k | 115k | 6 | 4 | 3/3 | 3/3 |
| t23 | Upload a file | 15 s | 12 s | 149k | 139k | 8 | 5 | 3/3 | 3/3 |
How it was measured
Setup
- Same agent harness (Codex CLI), same AI model and reasoning setting, same machine. Only the browser tool differs.
- Playwright MCP 0.0.83 in headless Chrome, with its coordinate tools enabled; axuse in its own browser.
- Both ran in a clean agent profile with no other browser or computer-use tools available, so neither could borrow another tool.
- 23 tasks that browser agents commonly trip over: forms, wizards, tables, iframes, shadow DOM, infinite scroll, dialogs, new tabs, downloads, drag and drop, editors, comboboxes, canvas, uploads, a shop checkout and a virtual grid.
Scoring and honesty
- A run passes only if the test site's own server records the right result and the agent reports the right value. Runs that bypass the browser are disqualified.
- Time is wall-clock per task; tokens are the AI model's total token use for the task; a tool call is one request to the browser tool.
What this does not measure. The test pages are plain public sites with no login and no bot protection, so this compares speed, cost and accuracy on everyday browser work. It says nothing about axuse's other differences: using your own signed-in Chrome, handling verification challenges, and the workflow plug-ins. Playwright MCP is also not slow here; the gap is a median of 3 seconds and 24k tokens per task. Want the raw per-run data? Email [email protected].