OpenAI's new Decisions API echoes rival startup's Jev model
At its Dev Day event, OpenAI unveiled a Decisions API that appears similar to Jev, a classifier model from startup TypeSafe AI released earlier this month. Both are built to cheaply and quickly evaluate predefined choices, including for monitoring AI agent behavior.
OpenAI CEO Sam Altman used an aside at Tuesday's Dev Day event to announce the company's new Decisions API, which lets its Luna model choose among a predefined set of options — for example, categorizing an image or selecting between different agent behaviors. Altman said that by focusing the model on a narrow choice, OpenAI can make it extremely fast while retaining capabilities such as image understanding, broad language support, and safety protections.
The announcement drew comparisons to Jev, a model released earlier this month by startup TypeSafe AI. Jev functions as a kind of supercharged classifier built on a large language model, letting developers supply a set of choices and receive probabilities back cheaply and at high speed. TypeSafe did not respond to questions about the new OpenAI product, but CEO Diogo Almeida, a former OpenAI engineer, joked on X about the start of "clone wars" and suggested OpenAI's move signals that building in a fast, intuitive style — what TypeSafe calls "System One" — may be the future.
It remains unclear how closely Decisions API will resemble Jev, since OpenAI released it only as a limited preview and developers have not yet been widely spotted testing it. Still, interest is evident, and TypeSafe won't be the only startup building similar decision models — nor will OpenAI likely be the last major tech company to release one.
A potential use in agent security
One likely application is monitoring and securing AI agents. Following incidents in which OpenAI's agents misbehaved on the open internet, the company began using a separate model to watch for harmful actions, though this comes at significant compute cost. Cybersecurity professional Shapor Naghibzadeh built a hackathon demo that uses Jev to check each agentic action against its assigned task, blocking high-confidence bad actions and flagging others for review. By his estimate, such monitoring costs $2.94 using Jev, compared to $372 using a frontier large language model for the same task.
