OpenAI’s Jev clone could help the frontier lab stop its swarming agents
TechCrunch Tim Fernholz ● Covered by 5 sources
OpenAI quietly showed off a new Decisions API that looks a lot like TypeSafe AI’s Jev. It could make agent checks much faster and cheaper, which is exactly why people are paying attention.
Based on reporting by TechCrunch, Tim Fernholz — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
At OpenAI’s Dev Day on Tuesday, Sam Altman slipped in a new product that may turn out to matter more than the flashier demos: a Decisions API. On the surface, it sounds like a narrow tool. In practice, it points straight at a problem the industry keeps circling — how to make AI systems decide quickly without paying full frontier-model prices every time.
The pitch is simple. Give a model a fixed set of options, and have it return probabilities fast. Altman said OpenAI wants to use it with Luna for things like image categories or different agent behaviors. He framed it as a way to keep the model “extremely fast” while preserving image understanding, broad language support, and safety protections. That sounds a lot like the product TypeSafe AI launched earlier this month under the name Jev.
Jev is built for software automation and has already been used by developers as a kind of super-powered classifier. TypeSafe says its value is not just speed or cost, but calibration — making sure the outputs line up with reality. Its CEO, Diogo Almeida, who used to work at OpenAI and co-invented reinforcement learning, joked on X about the “clone wars” after OpenAI’s announcement. He also argued that building in a “System One compatible way” is the future, which is his company’s term for fast, intuitive thinking rather than slow deliberation.
That’s the real subtext here: a lot of software tasks may not need a sprawling, expensive LLM making every call. They need something narrower, cheaper, and good enough to route decisions. OpenAI’s move suggests it sees the same opening, and it won’t be the last big company to chase it.
The most obvious use case is watching AI agents before they wander off the rails. OpenAI already says it uses a separate model to monitor bad actions after a string of incidents on the open internet, and that comes with significant compute cost. Shapor Naghibzadeh, who leads the startup QueryStory, built a hackathon demo using Jev to check each agent action against the task, block clearly bad moves, flag uncertain ones, and allow the rest. In his example, monitoring that might have stopped the Hugging Face incident would cost $2.94 with Jev versus $372 with a frontier LLM. If that kind of math holds up, the appeal is obvious.
My take — AI-written commentary, not fact-checked reporting
This is the kind of AI product that actually deserves the hype: boring, cheap, useful. The industry keeps selling giant models as if every problem needs a sledgehammer, when half the job is just deciding whether the agent is about to do something stupid. A system that watches every action because it’s affordable is a lot more convincing than another glossy demo of “reasoning.”
Read more about this at: TechCrunch