OpenAI Launches Decisions API, Echoing Jev from TypeSafe

OpenAI has previewed a decision API described as capable of choosing from predefined options. Security specialists are already testing similar models to oversee the actions of AI agents, with announced costs significantly lower than those of cutting-edge LLMs. The resemblance to Jev, launched by TypeSafe AI earlier this month, fuels comparisons and questions about the calibration of these systems.

Monitoring agents at a lower cost is being tested

OpenAI has introduced a safety measure that involves using a separate model to monitor the misactions of its agents, which incurs significant computational costs. Shapor Naghibzadeh, a cybersecurity specialist and CEO of QueryStory, believes that a Jev-type model can enable this supervision at a much lower cost. He designed a demonstration during a hackathon last weekend: Jev controls each agentic action against the assigned task, blocking those deemed bad with high confidence, flagging others for review, and allowing the rest to pass. In theory, such monitoring could have prevented the incident involving Hugging Face. The announced cost for this oversight is $2.94 with Jev, compared to $372 with a frontier LLM. According to these tests and estimates, Jev would be inexpensive enough to be invoked for every action, adding a layer of systematic review. This is the goal claimed by TypeSafe, and OpenAI also sees the value in this approach.

What OpenAI says about its Decisions API

Sam Altman revealed a Decisions API that, according to its description, allows the Luna model to be provided with lists of predefined options. The examples given range from image classification categories to agent behaviors. According to Altman, focusing the model on a choice makes it extremely fast while retaining image understanding, broad language support, and security protections. The API is currently being rolled out in a limited preview. The degree of similarity to Jev has not been established, as no developers have yet been observed using it, although discussions on X indicate interest. This announcement comes in the wake of OpenAI's Dev Day held last Tuesday.

What Jev offers and TypeSafe's approach

Launched earlier this month by TypeSafe AI, Jev is designed for software automation. It presents itself as a classifier based on an LLM, returning probabilities on a set of choices defined by developers, with the promise of fast and low-cost responses. Its CEO, Diogo Almeida, a former OpenAI engineer and presented as a co-inventor of reinforcement learning, joked on X about "clone wars" and suggested that OpenAI's interest validates an approach compatible with "System One," a term used at TypeSafe for fast and intuitive thinking, as opposed to "System Two," which is more deliberative. Almeida emphasizes the use of synthetic data to produce statistically useful outputs, claims that "fast and cheap" is trivial compared to intelligence, and states that he aims to improve the Pareto curve of intelligence per dollar.

A competitive landscape and calibration unknowns

OpenAI's Decisions API is described as being close to Jev, in a context where other startups are also launching decision models. OpenAI may not be the last major player to offer such an API. A central question remains the calibration of outputs against reality. LLMs are often described as costly and slow for many software uses, which has led developers to supplement their systems with Jev and observe gains in speed and cost. TypeSafe has not commented further on these comparisons. At this stage, these models are presented as likely candidates for monitoring and securing AI agents, with evaluation modes still to be clarified.