Jev 发布两周后 OpenAI 推出 Decisions API
视频梳理时间线：TypeSafe 的 Jev 九月中发布引爆开发者圈，两周后 OpenAI 在 DevDay 拿出功能相近的 Decisions API，作者借此讨论「先发只值两周」和「够用就好」的现实。
Mid-September: an ex-OpenAI researcher launches Jev and dev Twitter loses its mind. September 29: OpenAI ships its own version at DevDay. About two weeks. That's how long a hyped AI startup gets now. Quick context if you missed it. Jev, from Diogo Almeida's TypeSafe AI, is a "decision model." You give it a question, a fixed list of allowed answers and some context, and it returns one choice your code can branch on. Route this ticket to billing or support. Pick the agent's next tool. Flag this or let it through. Anyone who has glued together "respond ONLY with one of these words" + a JSON schema + a parser + a retry loop understood the appeal immediately. OpenAI's answer is the Decisions API. It works on the same idea: predefined answers, text or image context, and use cases in classification, routing and agent next steps. It's in limited preview now, with broad release promised in the coming days. Reported latency is around 150 ms, roughly 10x faster than a normal GPT-6 Luna call (launch claims, not benchmarks). It's not a straight copy. Jev is trained specifically for decisions, while OpenAI took Luna, its cheap model, and constrained the output. That's the uncomfortable part. OpenAI didn't train a new model to respond. They fenced in one they already had and shipped it as an endpoint. What I take from it: 1/ Being first buys you weeks. Jev had the hype and the demos, and got about 14 days before a comparable product was on the DevDay stage. 2/ Going viral is free market research for the incumbent. You prove demand and show which features people care about. The lab best placed to copy you reads the same timeline you do. 3/ "Better" often loses to "already on the invoice." Plenty of teams will take the good-enough version inside the OpenAI account they already pay for over a new vendor, a new security review and a new contract. 4/ The model is rarely the moat. If your product is "a model that does X," a lab can point an existing model at X. The harder things to copy are domain data, evals customers trust, calibration on real traffic, deep workflow integration, and being the neutral option for teams avoiding lock-in. 5/ Go narrower than they will. Decisions API was one of 20+ major DevDay announcements. Horizontal building blocks are exactly what labs ship. The messier and more industry-specific your problem, the less likely it becomes someone's keynote bullet. If you're building on a frontier lab's API today: which part of your product would take them longest to rebuild? Be honest.