Dax benchmark, whole job
Seconds, lower is better. Grey: the public ComputeSDK leaderboard, top 5. Violet: our own runs with the same script; we are not on the leaderboard yet.
Show as a table
| Option | s | Shape |
|---|---|---|
| Isorun | 34.5 | Leaderboard, 9 October 2026 |
| Namespace | 40.1 | Leaderboard, 9 October 2026 |
| Uno USA, High‑end · our run | 40.6 | Our run through our public API, median of 5, range 39.6–41.8 s. Not on the leaderboard yet |
| Gravix | 41.7 | Leaderboard, 9 October 2026 |
| GiveMeANode | 41.7 | Leaderboard, 9 October 2026 |
| Blaxel | 45.9 | Leaderboard, 9 October 2026 |
| Uno Netherlands · our run | 53.4 | Our run, 7 October 2026 |
Leaderboard of 9 October 2026, 21 providers, one run per provider. Our numbers are medians of 5 runs of the same script on the same size. The same US server with no virtualization does it in 32.2 s: that is the ceiling, and the microVM that keeps your computer apart from its neighbours costs about six seconds of it.
What Dax measures
Dax times the work an agent does before it can start coding: set up the toolchain, clone a repository, install about 950 MB of packages and typecheck the project. It runs on a computer created for that run, so nothing is left warm from the last one.
Two phases make most of the time. Typecheck is the processor, so the newer chip in the US wins there. Install is the network: the packages arrive at full port speed, and a faster port is our next step.
What it means if you do not write code
An agent spends part of every task installing, building and checking. On High‑end that wait is about 24% shorter, and it adds up over a day of work.
Two regions: the closest one, or the fastest
Uno suggests the region closest to you. In the US, computers run on High‑end servers with our fastest processors.
Netherlands
StandardClosest to Europe, Asia and Africa
- NVMe disks
- DDR5 with error correction (ECC)
- Dax: 53.4 s
USA
High‑endClosest to the US, Canada and Latin America
- NVMe disks, our fastest processors
- DDR5 memory
- Dax: about 40 s
The same plans in both: what changes is the price and the speed. You pick the region when you create a computer, by API too; with no region given, your account's region is used. On Max and Enterprise both regions come at no extra cost.
Geekbench: a second MacBook in the cloud
Geekbench 6 is the score most laptops are compared by. Measured in the Netherlands, 16 vCPU score like a MacBook Air M4, and 8 vCPU like a MacBook Air M2.
Geekbench 6, multi-core
Higher is better. Uno computers against MacBook averages.
Show as a table
| Option | points | Shape |
|---|---|---|
| Uno, 16 vCPU (Pro) | 15,488 | Measured August 2026 |
| MacBook Air M4 | 14,924 | Geekbench Browser average |
| Uno, 8 vCPU | 10,449 | Measured August 2026, inside an Uno computer |
| MacBook Air M2 | 9,684 | Geekbench Browser average |
Results are public: 8 vCPU, 16 vCPU. Measured in August 2026 on an idle server.
How we measured
- Their script, unchanged. We ran the ComputeSDK Dax script as published, with no edits to the job.
- Their size. 8 vCPU · 16 GB, the size the public leaderboard uses.
- A fresh computer every run. A new Firecracker microVM with the same isolation our customers get.
- Timed from outside, through our public API, with the ComputeSDK adapter, the way their runner does it, so start-up counts.
- Medians of 5 runs, with the range shown. Every run finished.
We are not on the public leaderboard yet. We will submit, and this page will show our official result next to theirs.
What these numbers do not say
- The US numbers come from a server with no other customers on it. vCPU are shared, so with busy neighbours a run takes longer.
- Memory in the US region has no error correction (ECC). In the Netherlands it does.
- From Europe, a computer in the US answers your screen about 0.1 s slower. Agents and builds do not notice.
- Our runs already go through our public API, the way the leaderboard calls every provider. Their runner sits elsewhere, so the official time can differ a little from ours.
- A benchmark is one job. Your project can be faster or slower; the gap between the regions holds for work that leans on the processor.