OpenAI CEO Sam Altman used Dev Day on Tuesday to disclose Decisions API, a tool that lets the Luna model choose between a predefined set of options instead of generating open-ended text. The product closely resembles Jev, released earlier this month by TypeSafe AI as a fast classifier for software automation. The disclosure matters because agent oversight remains expensive, while one demo put monitoring cost at $2.94 with Jev against $372 with a frontier LLM.
How Decisions API mirrors the Jev approach
TypeSafe AI released Jev earlier this month as a super-powered classifier built on a large language model. Developers supply a set of choices, and the model returns them as probabilities at low cost and high speed. Altman described Decisions API in similar terms, citing categories for image classification or different agent behaviors as examples of preset options. The product is available as a limited preview, and TechCrunch reported that developers have not yet tested it publicly in depth.
Altman said focusing the model on a narrow choice makes it extremely fast while retaining image understanding, broad language support, and safety protections. That design trades open-ended generation for constrained selection, which suits repetitive software decisions. TypeSafe frames the same idea as System One compatible operation, meaning fast intuitive judgment rather than deliberate System Two reasoning. The shared logic is to reserve heavyweight models for complex work and route routine choices elsewhere.
The announcement fits a wider shift away from using general models for every software task because they are comparatively slow and expensive. Developers have used Jev to augment large language models and reported faster and cheaper operation, according to the report. Decisions API is not the only Jev-like offering, as other startups are releasing similar models and OpenAI is unlikely to be the last large technology company to follow. Discussion on X showed clear interest, although direct comparisons of accuracy and calibration are still missing.
What cheaper decision models mean for agent operations
For companies running agents, the practical effect is a lower-cost layer that can review or steer behavior at scale. TypeSafe CEO Diogo Almeida, a former OpenAI engineer credited in the report as a co-inventor of reinforcement learning, joked on X about the beginning of the clone wars. He said OpenAI interest suggests System One style building may become standard practice. Small teams gain access to classification, routing, and policy checks without paying frontier-model prices for each call.
The limits center on output quality rather than speed, because fast and cheap alone does not guarantee useful decisions. Almeida said his company relies on synthetic data to produce statistically useful outputs and described intelligence per dollar as his main objective. A central open question is how well calibrated each decision model will be to real conditions. Buyers should therefore test accuracy on their own categories, ask how calibration is measured, and avoid treating the preview as equivalent to Jev before independent results appear.
A concrete use case is agent monitoring, where a separate model watches for unwanted actions at significant compute cost after incidents involving misbehaving agents on the open internet. Shapor Naghibzadeh, a cybersecurity professional leading QueryStory, built a hackathon demo last weekend that checks each agentic action against its assigned task. The demo blocks actions judged bad with high confidence, flags uncertain cases for review, and permits the rest. Its author argues that such screening could have stopped the Hugging Face incident, with Jev cheap enough to run on every action.
