BenchmarksComing soon

Karpo BrowserBench

A benchmark for browser agents: given a real web page and a task, how well does a model decide the next step?

Real websites
Tasks are taken from real sites people use, not from pages built for a test.
Step by step
It scores each decision along the way, as well as whether the task ends where it should.
For decision models and agents
It is the yardstick we use for Karpo's own browser agent and the Karpo Decision models.

The tasks, the method and the first results will be published here.

Read the blog in the meantime