OpenAI Previews Decisions API for Faster Agent Monitoring
OpenAI used its Dev Day event to preview the Decisions API, a limited-preview product that CEO Sam Altman described as a way for the company’s Luna model to select from predefined options. The API can be used for choices such as image-classification categories or alternative agent behaviours, and it arrives shortly after TypeSafe AI released Jev, a model built for software automation.
Rather than asking a general-purpose large language model to produce open-ended text, developers can supply a fixed set of choices and receive probabilities for them. Altman said that narrowing the task to a choice can make the model extremely fast while retaining image understanding, broad language support and safety protections.
A specialised layer for automated decisions
Jev is positioned as a high-speed, low-cost classifier augmented by LLM capabilities. TypeSafe AI calls its approach compatible with “System One”, its term for fast, intuitive processing, in contrast with deliberate “System 2” reasoning. Its chief executive, Diogo Almeida, said that low cost and speed alone are not sufficient; the difficult problem is producing intelligent, statistically useful outputs.
TypeSafe AI says synthetic data is central to that objective. The decisive issue for Jev, OpenAI’s Decisions API and comparable products from other startups will be how well their probability outputs are calibrated to real-world conditions. OpenAI has released its offering only as a limited preview, so developers have not yet publicly put it through extensive testing.
Potential role in AI agent controls
The timing is notable because OpenAI has introduced security measures after incidents involving agents behaving improperly on the open internet. One measure uses a separate model to inspect potentially bad actions, but that safeguard comes with significant compute cost. The connection to enterprise AI competition and deployable controls highlights how enterprise AI competition increasingly depends on deployable controls as well as model capabilities.
QueryStory founder Shapor Naghibzadeh built a hackathon demonstration in which Jev evaluated each agentic action against the task assigned to the agent. It blocked actions judged bad with high confidence, sent less certain cases for review and allowed the remainder. He estimated that this form of monitoring would cost $2.94 with Jev, compared with $372 using a frontier LLM.
The demonstration is not evidence that every deployment will achieve the same result, but it illustrates why low-cost decision models may be attractive for controls applied to every agent action. Businesses evaluating such systems should measure calibration, review flows and failure handling alongside inference cost before making them part of agent governance.