OpenAI’s ‘Jev Clone’ Could Help Monitor AI Agents
OpenAI has unveiled its new "Decisions API," a tool designed for rapid, cost-effective decision-making, drawing comparisons to TypeSafe AI's Jev model. This innovation addresses the limitations of traditional LLMs for certain software tasks and holds significant promise for applications like enhancing AI agent monitoring and security. The competitive landscape for these efficient decision models is rapidly evolving.
OpenAI’s new Decisions API, unveiled during its DevDay 2026 event, appears to offer functionality similar to Jev, a model launched earlier this month by TypeSafe AI for software automation. Jev works as a specialised decision-making model, allowing developers to define a set of possible choices while the model returns probabilities rather than generating free-form text, making it faster and cheaper for certain tasks.
OpenAI’s Decisions API similarly uses Luna to classify inputs, route requests or select actions from predefined answers, with CEO Sam Altman saying that constraining the model to specific choices can make it extremely fast while retaining capabilities such as image understanding, multilingual support and safety protections.
The announcement drew a playful response from TypeSafe AI CEO Diogo Almeida, a former OpenAI engineer and co-inventor of reinforcement learning, who referred on X to the beginning of “clone wars.” Almeida has described TypeSafe’s approach as “System One,” referring to fast, intuitive decision-making as opposed to the deliberate reasoning associated with System Two, while arguing that conventional LLMs can be unnecessarily slow and expensive for some software tasks.
However, the extent to which Decisions API is technically comparable to Jev remains unclear, as OpenAI has released the product in limited preview and extensive independent developer testing has not yet emerged; other startups are also developing similar decision-making models, making calibration and accuracy important measures of their usefulness.
One potential application is monitoring AI agents, particularly as companies look for cheaper ways to review increasingly autonomous systems. Cybersecurity professional Shapor Naghibzadeh, who leads QueryStory, built a hackathon demonstration using Jev to assess each agentic action against its assigned task, blocking actions judged highly likely to be malicious, flagging uncertain cases for review and allowing the rest; according to TechCrunch, the demonstration estimated monitoring costs of $2.94 with Jev compared with $372 using a frontier LLM.
The figures come from that demonstration rather than an OpenAI benchmark, but they illustrate the central proposition behind these smaller decision models: if sufficiently accurate, their low cost could make continuous checks of AI-agent actions more practical as developers build systems that operate with greater autonomy.