Jev vs Decisions API: both perfect, Jev faster in 9 of 10

The comparison comes as TypeSafe's Jev adds image input, letting its System One requests carry pictures alongside text. The author sets up the test to see what Jev's image handling can do and how it stacks up against OpenAI's Decisions API, which the author uses as the reference. Jev's image and Box features are still positioned for prototype use, with request formats, limits, and accuracy subject to change between releases.

The test covers five task types: comparing to a known photo, classifying by text criteria, classifying by example images, spotting change across frames, and locating an object. Jev and Decisions API handle these differently. Jev can define choice options with example images, and its Box feature returns positions; Decisions API has no parameter for image-defined choices, so the author labels images and places them in the same message. That difference in how images are supplied is part of why request sizes and timings diverge.

The accuracy results were a clean sweep for both, which the author attributes to using photos that are easy to judge by eye. The more telling differences were in confidence values: Decisions API tends to return values close to 0 or 1, while Jev returns more moderate ones like 0.93 for the same-dog case; for the gesture, Jev returned confidence 1.00 while Decisions API left a 0.04 on fist. None landed near the 0.5 threshold. On speed, Jev's total wait was 3.639 seconds versus 6.269 for Decisions API, though the author notes these times include image upload and are single measurements, so small gaps could flip. On cost, Jev totaled about 0.4 cents versus about 0.26 cents, a small difference at this scale that would grow with call volume.