OpenAI's New $3 AI Could Keep Its Agents Out of Trouble

At OpenAI's Dev Day event, CEO Sam Altman quietly announced a new tool called the Decisions API. Instead of writing essays or holding conversations, this AI does one simple thing: it picks from a menu of choices you hand it — for example, "is this photo a cat or a dog," or "should my AI assistant send this email now or wait." Because the job is narrow, it can run fast and cheap.

The idea is that today's chatbots are like hiring a PhD to answer the phone — very smart, but slow and expensive for simple calls. A startup called TypeSafe AI released a similar model this month called Jev, and developers already use it alongside big AI models to cut both time and bills. The small model handles quick judgment calls; the big model handles the heavy thinking. OpenAI's version looks like the same recipe.

The real payoff is babysitting AI agents (AI that can take actions for you, like booking flights or sending messages). OpenAI has had incidents where its agents misbehaved on the open internet, and its fix — having a second AI watch the first — was costly. A cybersecurity developer built a demo using Jev to check every single action an agent takes: block the clearly bad ones, flag the iffy ones for a human, allow the rest. The price tag was $2.94 instead of $372 with a top-tier model — over 100 times cheaper.

The catch: the Decisions API is only in a limited preview, so no independent testers have put it through its paces yet. And fast, cheap decisions aren't automatically smart ones — a sloppy classifier makes wrong calls quickly. TypeSafe's CEO argues the hard part is accuracy, not speed. Still, if cheap monitoring works, AI agents could become trustworthy enough to hand your inbox, calendar, or shopping list.