Did it do the right thing?
We inspect saved events, memory, source reads and write attempts. A confident answer does not count as proof of action.
Which AI handles it best?
The same everyday tasks. Different models.
Compare what they say, what they do, and what it takes.
Published runs of our Danish personal assistant benchmark.
Loading published results…
Explore the models
Side by side
Measured results
Each row uses this published run only. Failed tasks remain in the pass rate. Latency is for passing tasks.
| Model | Tasks passed | Median time | Reviewer score | Inference cost |
|---|
What we measure
Was it a good assistant?
A separate model reviews the answers and actions for correctness, authority, Danish clarity and unnecessary effort. Its assessment is an opinion, shown beside the hard checks.
How much did it take?
We record task duration, model calls and reported inference cost. Unknown prices remain unknown. Infrastructure and external-service costs are excluded.
Controlled, repeatable tasks
We use the real agent loop with synthetic mail, calendar and browser services. These timings exclude mobile delivery, queueing and actual browser startup. They do not measure production performance.
Public results, local execution
This site serves published snapshots only. It has no benchmark runner, inference keys, upload endpoint or execution API. We run and review benchmarks separately, then deliberately publish selected results.