At OpenAI’s recent Dev Day, CEO Sam Altman announced a limited preview of a Decisions API. The service is described as a classifier built on a large language model that can receive a list of options and return probability scores. According to Altman, the design mirrors a model called Jev that TypeSafe AI released earlier in the month, aiming to deliver rapid, low-cost decision making for software tasks.
Altman explained that the API will be used to steer OpenAI’s Luna model by presenting it with a predefined set of alternatives, such as image categories or possible agent behaviors. By constraining the model’s focus to a specific choice set, the system can operate at higher speed while preserving capabilities like image comprehension, broad language understanding, and built-in safety layers.
TypeSafe AI’s Jev model, launched just weeks before the OpenAI announcement, was created to accelerate software automation. The company’s chief executive, Diogo Almeida,who previously worked at OpenAI and helped develop reinforcement learning,responded on X with a light-hearted comment about “clone wars.” TypeSafe declined to provide further comment to TechCrunch about the new OpenAI product.
The emergence of Jev-style decision services is prompting several startups to introduce comparable offerings. A central concern among observers is how accurately these models will map their probability outputs to real-world outcomes. Almeida highlighted that TypeSafe’s advantage lies in synthetic data generation, noting that “fast and cheap is very easy… intelligence is the hard part,” a statement reported by TechCrunch.
OpenAI has also begun deploying a separate model to watch for undesirable actions by its autonomous agents, a measure introduced after a series of misbehaviors on the open internet. Shapor Naghibzadeh, who leads the cybersecurity firm QueryStory, built a hackathon demo that uses Jev to evaluate each agentic step, blocking actions with high confidence of being harmful and flagging ambiguous cases for review.
Cost analysis presented by Naghibzadeh shows that running Jev for monitoring costs roughly $2.94 per instance, compared with $372 when using a frontier-class large language model. This price gap suggests that Jev-like tools could be deployed on every agentic decision, providing a continuous safety layer that may improve overall reliability of autonomous systems, an outcome both TypeSafe and OpenAI appear to value.