OpenAI Made a Jev Competitor, So I Tested It Against Jev and Clef

OpenAI just shipped the Decisions API with GPT-6 Luna, a decision ("System One") model that answers yes/no, picks from choices, or scores against a rubric, instead of writing text. I put it head to head with Jev (TypeSafe AI) and Clef / Clef Flash (Cloudflare) on text routing, tool-call checks, and vision tasks like find-on-screen, counting, charts and receipts. The short version: Luna had the fastest median on both (149 ms on text, ~220 ms on images), but it wasn't the most accurate: 93% vs Je