Benchmarks & research
QApilot’s screen-classification study found Jev cheaper and faster than GPT-5.1, but substantially less accurate on its 371 mobile-app screens.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Bonsai describes the screenshots in text before Jev sees them, but that step’s latency is absent from the timings. The maker reports 59.8% accuracy for Jev against GPT-5.1’s 93%; the ambiguous “others” category caused trouble. Different average confidence on right and wrong answers does not, by itself, make a reliable error detector.
The first AI reads an image and describes the screen. Jev then uses that text to choose a screen category. If you adapt this two-model approach, measure the entire chain: the description step can add both time and mistakes, which a test of Jev alone would miss.