TypeSafe 的 Jev 与 OpenAI Decisions API 的 14 天竞速

Aakash Gupta @aakashgupta原推发布于 2026.10.07

在 X 上打开原推

An ex-OpenAI researcher spent two years in stealth building a new category of AI model. 14 days after he launched it, OpenAI announced its own version. Today it went to public beta for every developer. Diogo Almeida left OpenAI to found TypeSafe, raised $40M, and shipped Jev on September 15. A model that doesn't generate text at all. You send it application state and typed questions, it returns decisions with calibrated probabilities. 70 to 500 milliseconds, $42 per billion input tokens, output free. Then at DevDay on September 29, OpenAI announced the Decisions API. Questions in, predefined answers with confidence scores out, 150 milliseconds, powered by GPT-6 Luna. You cannot build a new API product in two weeks. OpenAI was already building this when Jev launched, which means the startup and the incumbent were racing toward the same category in parallel and the startup won by exactly 14 days. For years the advice to founders was don't build a wrapper, build a real model. TypeSafe did everything that playbook asks. New architecture, new training method called RLCD, a genuinely new model class. OpenAI matched the category anyway, inside a month, with millions of developers already holding API keys. Watch the pricing. TypeSafe charges $0.042 per million input tokens and nothing for output. OpenAI hasn't published a Decisions API price at all. When an incumbent clones your product and won't name a price, the price is the entire fight still to come. And the category itself is bigger than it looks. Routing, classification, guardrails, and agent action checks run thousands of times for every one user-facing response. Whoever owns the decision layer sits inside every loop of every agent. Fourteen days. That's how long a brand new model category stayed a moat.

这条是引用着另一条写的：x.com/1633874951508721686/status/210757338222918…