Measured on real tasks
17 of 18 research answers12 of 12 right tool picks8 of 8 launch films3 of 3 landing pages
Web research17 of 18
17 Oct 2026
Web research
Search Star answered 17 of 18 timed research questions, with a median of 6.5 seconds.
The right tool12 of 12
21 Oct 2026
The right tool
Desktop Browser picked the right tool for 12 of 12 everyday tasks, twice in a row.
Launch films8 of 8
20 Oct 2026
Launch films
Product Videos finished 8 of 8 launch films, each in under 3 minutes.
Landing page and logo3 of 3
9 Oct 2026
Landing page and logo
Company Design built a landing page and its logo 3 of 3 times, each in under 9 minutes.
How we test
Every number here comes from our own tests of real tasks, run in October 2026.
Real tasksEach test is work a person would hand over, like a bug or a research question.
Hidden checksA checker the agent never sees decides if the work passed.
Every resultWe show the counts as they came out, small samples included.
Frequently asked questions
Are these public benchmarks?
No. They are our own tests on real tasks. Each one is small, so we show counts, not percentages.
Why do the dates differ?
Each test runs when its part of Starteammate changes. We show the latest run of each.
Does an agent ever fail a task?
Yes. The misses are in the counts above. We use them to decide what to fix next.
