26 tasks.
4.2 seconds each.
Ultrafast passed every task in our browser-use harness, three full runs in a row, at a 4.2–4.3 second average. That makes it the fastest browser use ever made. The same task set takes 7.8 seconds in Chat mode. Below is every task, every time, and the method, so you can rerun it yourself.
Fast against Chat
One number for the whole loop.
Fast and Chat run the same browser, the same tools and the same task pages. The difference is how a decision is made: one operation-and-element choice per step, instead of a full tool-calling conversation.
| Measure | Fast (ultrafast) | Chat |
|---|---|---|
| Success | 26/26 (100%) on three full runs | 17/17 on the original task set |
| Average time per task | 4.2–4.3s | 7.8s |
| Four hardest multi-step tasks | 4/4 · 6.2s average | 4/4 · 22.6s average (one took 71s) |
| False “done” claims | 0 | 0 |
| Decision model | Mercury Decide (Inception) | Mercury 2 |
One full Fast run of 26 tasks: 69 browser actions, 127 decision calls and 4 text-writing calls. Chat was measured on the original 17-task set, so read the averages next to that difference.
How it got there
Four upgrades, one loop.
The score did not start at 26/26. It started at 10/17 with one false completion. Each line is a change to the loop, not to the task set.
| Change | Score | Note | Avg time |
|---|---|---|---|
| Before the upgrades | 10/17 (59%) | 1 false “done” | 3.2s |
| Better page reading and new actions | 17/17 | — | 3.9s |
| Checklist and chat-model fallback | 21/21 | — | 4.4s |
| Answers to questions | 26/26 | — | 4.2–4.6s |
| Text extracted straight from the goal | 26/26 | — | 4.2s |
Every task
The full 26, in order.
Times include the model calls, the page reads and the action. A task with 0 actions is a question answered from the page without changing anything.
| Task | Actions | Time |
|---|---|---|
| Add a specific product to cart (6 identical buttons) | 1 | 1.6s |
| Cheapest item to cart | 1 | 1.8s |
| Answer: price of a product | 0 | 2.1s |
| Answer: cheapest product + price | 0 | 2.4s |
| Answer: “not on the page” (no made-up answer) | 0 | 2.5s |
| Close a dialog with Esc, then click | 2 | 2.6s |
| Link that opens a new tab, then confirm | 2 | 2.6s |
| Search with no button (press Enter) | 2 | 2.9s |
| Hover menu → submenu item | 2 | 2.9s |
| Custom dropdown menu sort | 2 | 3.3s |
| Search + count results (answer) | 2 | 3.4s |
| Pagination to item 27 | 3 | 3.7s |
| Tabs widget → mark review helpful | 2 | 3.7s |
| Date picker + continue | 2 | 3.8s |
| Brand + price filters | 3 | 3.9s |
| Autocomplete pick (Zurich) | 3 | 4.4s |
| Unlabeled search box | 2 | 4.4s |
| Button far below the fold | 5 | 5.0s |
| Flight form + checkbox | 3 | 5.2s |
| Cookie wall, then subscribe | 3 | 5.7s |
| Search + sort by price | 3 | 5.9s |
| Search + sort + in-stock filter | 4 | 6.1s |
| Scroll inside a panel to find Kyoto | 8 | 6.2s |
| Form inside a dialog (address) | 4 | 6.9s |
| Answer: phone number in footer | 4 | 7.0s |
| Two autocompletes + travellers (trip search) | 6 | 8.5s |
Real sites
16 live sites. 16 passed.
The same Fast loop on the open web: Wikipedia, GitHub, YouTube, MDN, Hacker News and four form-test sites, run through the MCP server with run_task. 16 of 16 passed at an 8.2s average. Fast finished 14 on its own; on the other 2 the chat model took over where Fast got stuck, without redoing the finished steps.
| Site | Goal | Result | Actions | Time |
|---|---|---|---|---|
| Wikipedia | “Search Wikipedia for Ada Lovelace and open her article” | article open | 2 | 7.8s |
| Wikipedia | “In which year was Alan Turing born?” | answered 1912 | 2 | 7.3s |
| Wikipedia | “Mount Everest elevation in metres” | answered 8,848.86 m | 2 | 7.6s |
| GitHub | “Search repositories for puppeteer and open puppeteer/puppeteer” | repo open | 4 | 9.1s |
| YouTube | “Search YouTube for lofi hip hop radio” | results page | 3 | 6.7s |
| MDN | “Search MDN for Array.prototype.map and open its page” | page open (chat takeover) | 5 | 20.5s |
| Hacker News | “Title of the top story right now” | answered | 0 | 2.4s |
| Books to Scrape | “Price of "A Light in the Attic"” | answered £51.77 | 0 | 1.4s |
| Books to Scrape | “Open Poetry, then the book "Olio"” | book open | 2 | 5.6s |
| DemoQA | “Fill name, email and address, then submit” | submitted | 4 | 9.5s |
| DemoQA | “Select the "Impressive" radio option” | selected | 1 | 1.7s |
| DemoQA | “Expand Home and tick Desktop” | ticked (chat takeover) | 4 | 35.0s |
| The Internet | “Choose Option 2 in the dropdown” | selected | 1 | 1.6s |
| The Internet | “Make sure both checkboxes are ticked” | ticked | 1 | 1.5s |
| The Internet | “Add three elements” | 3 added | 3 | 3.5s |
| TodoMVC | “Add two todos, mark one completed” | done | 6 | 9.3s |
Script: scripts/bench-real.mjs. Live sites change and sometimes show bot checks, so expect some run-to-run variation.
Live runs
Real sites, your Chrome.
The harness is not the whole story, so four goals were run on the live web in a normal Chrome profile. Four runs is a demonstration, not a statistic.
| Goal | Result | Time |
|---|---|---|
| “Open YouTube and search for trending videos of today” | results page | 6.7–9.4s |
| “Search Wikipedia for Gödel's incompleteness theorems and open the article” | article open | 10.2s |
| “Open Wikipedia and find out in which year Gödel published his incompleteness theorems” | answered 1931 | 14.4s |
| “Search Wikipedia for Ada Lovelace and find out which year she was born” | answered 1815 | 11.0s |
Method
Reproducible, not a slide.
The harness ships with the extension. It serves 26 local pages, drives each goal through
run_task in ultrafast mode, and scores the page state or the answer.
What is measured
- Task pages. 26 local pages with forms, dialogs, dropdowns, hover menus, pagination, tabs, cookie walls and a scrollable panel.
- Scoring. A task passes only if the page ends in the expected state, or the answer matches the page. A model saying “done” is not a pass.
- Model. Mercury Decide (Inception) for Fast, Mercury 2 for the Chat comparison. Both calls and page reads are inside the measured time.
- Runs. Three complete Fast runs of all 26 tasks. Times shown are from the latest complete run.
- Machine. One desktop machine, Chrome stable, ordinary network. No GPU and no special hosting.
Run it yourself
OPENROUTER_API_KEY=sk-or-... \ DECISION_MODEL=inception/mercury-decide \ node scripts/bench-fast.mjs
Set BENCH_MODE=normal to run the same tasks in Chat mode. The repository keeps
the harness at 26/26 for Fast-mode changes, so a regression shows up as a failed task, not a
slower average.
Why we call it the fastest browser use ever made
Because that is the claim the numbers support. 4.2 seconds per task, 26 of 26, no false completions, and a harness anyone can rerun. We have not audited every browser agent ever shipped, and we are not going to pretend otherwise. If you have a reproducible run that beats this, send it and we will link it here.
What this benchmark is not
- It is not a third-party audit. It is our harness, run on our machine, published in full.
- The 26 pages are local and simpler than the live web: no bot walls, no A/B tests, no slow third-party scripts.
- Four live runs show the loop works outside the harness. They are not a sample size.
- Provider latency moves the average. A slower model produces slower runs with the same task score.
- “Fastest” here means fastest measured loop across the open-source browser agents we could test. It is not a claim about every proprietary system ever shipped.