Monitoring a misbehaving AI agent with a frontier LLM costs $372 per run. With Jev, TypeSafe’s fast-classifier model, that same job costs $2.94. That price gap is why OpenAI just built something that looks a lot like it.
At OpenAI’s Dev Day, CEO Sam Altman casually mentioned a new product called the Decisions API. As TechCrunch reported, the API gives OpenAI’s Luna model a predefined set of options to pick between, whether that’s image categories, agent behaviors, or classification tasks. By narrowing the model’s focus to a fixed set of outputs, OpenAI says it can deliver fast, cheap decisions without stripping out image understanding or safety protections. That’s almost exactly what Jev does.
TypeSafe AI released Jev earlier this month. It’s a classifier built on an LLM, designed to return probability scores across a set of choices quickly and at low cost. TypeSafe CEO Diogo Almeida, a former OpenAI engineer, responded to the news with a joke on X about the clone wars starting. But behind the humor is a real point: OpenAI copying your product concept is a strong signal that the concept works.
The bigger issue Jev and Decisions API are both responding to is that standard LLMs are often the wrong tool for software automation. They’re slow and expensive when you need fast, high-volume decisions. Developers have been layering Jev on top of LLMs as a cheaper routing and classification layer, and getting meaningful performance gains. This is what TypeSafe calls “System One” thinking, fast and intuitive, versus the slower deliberate reasoning of full LLM calls.
The most compelling use case right now is agent monitoring. OpenAI has had public incidents where its agents did things they weren’t supposed to on the open internet, and its current fix involves running a separate model to watch for bad actions. That works, but it’s expensive. A developer named Shapor Naghibzadeh built a hackathon demo using Jev that checks every agentic action against the original task, blocks high-confidence violations, flags edge cases, and clears the rest. The cost at Jev’s pricing: $2.94. The Hugging Face incident, often cited as a cautionary example, could theoretically have been caught by a system like this.
That’s the practical argument for this entire category of model. If decision-layer inference is cheap enough to run on every single agent action, it becomes a viable safety net rather than an occasional audit. No other approach gets you there at current frontier LLM prices.
Decisions API is still in limited preview and hasn’t been stress-tested publicly yet. And TypeSafe isn’t the only company in this space. Other startups are shipping similar classifiers, and more big labs will follow OpenAI. Almeida’s argument is that his moat is the synthetic data TypeSafe uses to keep outputs statistically well-calibrated to real-world conditions. “Fast and cheap is very easy,” he told TechCrunch. “Use dice. Intelligence is the hard part.” That’s the right framing. Speed and cost are table stakes. Calibration is where this category actually gets decided.
