OpenAI’s Decisions API could help monitor AI agents at lower cost

OpenAI introduced its Decisions API at Dev Day, where CEO Sam Altman described it as a way for the company’s Luna model to choose among predefined options, such as image categories or agent behaviors. He said focusing a model on a choice could make it faster while retaining image understanding, broad language support and safety protections.

The API appears to offer functionality similar to Jev, a model from TypeSafe AI designed for software automation. Jev returns probabilities for a set of choices, aiming to do so quickly and cheaply. OpenAI released Decisions API as a limited preview, and its similarity to Jev remains unclear.

TypeSafe CEO Diogo Almeida told TechCrunch that the difficult part is delivering useful intelligence, rather than speed and low cost alone. He said TypeSafe uses synthetic data to produce statistically useful outputs. Other startups are also developing similar APIs.

One possible use is checking AI agents’ actions. QueryStory CEO Shapor Naghibzadeh built a hackathon demo using Jev to assess each action against an agent’s assigned task: blocking actions judged highly likely to be bad, flagging others for review and allowing the rest. The source reports that monitoring in the demo cost $2.94 with Jev, compared with $372 using a frontier LLM. The article says OpenAI has separately used a model to watch agents for bad actions, at “significant compute cost.”