Benchmarks
Measured on a real machine, on real tasks
Every figure is a before-and-after on the same request, from Helix's own per-step timing log and live runs against the real model on one Windows 11 desktop.
Seconds to finish a task
The same requests, before and after this autumn's work. Grey is before, mint is now.
Measured this release
Each model call takes 3 to 7 seconds, so the speed comes from needing fewer of them.
| A learned task | about 1 s, no model call |
| Find a file by name or content | 0.04 to 0.2 s |
| Helix Code, first words on screen | 1.4 s |
| In-app update, click to restarted | 24 s, 22 MB patch |
+73%
more of a web page read
shadow DOM, on developer.mozilla.org
83%
smaller updates
a 22 MB patch, not the full installer
17/17
live tasks succeeded
desktop, web, files and media
1098
automated tests passing
the full suite, every release
Comparison
Against a typical cloud computer-use agent
Not a score: a capability table. Cloud agents run in a fresh VM somewhere else; Helix runs on the computer you are sitting at.
| Helix | Cloud computer-use agent | |
|---|---|---|
| Where it runs | Your own PC, with your apps and logins | A fresh VM in a datacentre |
| Your installed software | Whatever you have: Spotify, Office, launchers | Only what the sandbox image ships |
| Signed-in sessions | Already signed in: it is your browser | Needs credentials handed to it |
| Web work | Off-screen tab, by element, through DevTools | Screenshot plus coordinates, typically |
| Keep using the machine? | Yes for web, files and API work | Yes, it is not your machine |
| Permission model | Typed risk classes; asks before anything that commits | Varies; often all-or-nothing |
| Repeats what worked | Learned shortcuts, re-checked every time | Replans from scratch |
There are no head-to-head scores against other products here. Helix has not been run on OSWorld or WebArena, and invented numbers next to another company's name would be worthless to you and unfair to them.
Run the numbers yourself.
Every timing on this page comes from Helix's own log, which you can read on your machine.