TypeSafe AI's Jev judges only, beats GPT-6 Luna in 78-headline test

Jev arrives at a moment when generative AI competition is intense, but it deliberately does not play that game. Instead of producing text or images, it is limited to deciding — Yes/No, Score, or Choice — and its appeal is speed on large batches of small judgments. The test with AI Watch headlines illustrates the intended use: when a reader scans dozens of article titles, the decision of what to read now is exactly the kind of call Jev is built for. Prompts can be pre-set with the user's priorities to sharpen those judgments.

The comparison with GPT-6 Luna, however, shows the limits of that advantage. Jev responded to each decision in 100–500 milliseconds and completed 78 items in just over 40 seconds, while per-item calls to GPT-6 Luna took roughly ten times as long, totaling nearly 8 minutes. But when the 78 headlines were bundled into one GPT-6 Luna call, the total dropped to about 45 seconds — close enough that Jev's edge is not automatic. Cost was also modest on both sides, with one sample run costing less than $0.01 and GPT-6 Luna around $0.01 per call. The lesson the body draws is that model and method should be matched to the task.