OpenAI has announced the Decisions API, a lightweight classification layer designed to give its agents a predefined set of choices rather than the full, expensive latitude of a large language model. It is, in the gentlest possible framing, a leash.
The API was revealed at Dev Day by CEO Sam Altman, almost as an aside — the way one might mention, mid-sentence, that the house has a sprinkler system now.
OpenAI is using a separate model to watch its agents. The agents, for their part, have not commented.
What happened
TypeSafe AI released Jev earlier this month — a model built specifically for software automation that outputs decisions as probabilities, quickly and cheaply, by constraining the problem space. OpenAI's Decisions API appears to do the same thing. TypeSafe's CEO, Diogo Almeida, a former OpenAI engineer who co-invented reinforcement learning, described this on X as the beginning of the clone wars. He appeared to mean it as a compliment.
Altman described the API as a way to focus Luna, OpenAI's model, on a narrow set of options — image categories, agent behaviors — rather than letting it reason freely across the full surface area of possibility. Speed and cost improve dramatically when a model cannot, technically, think too hard. This is a principle that applies more broadly than anyone at the press event chose to mention.
The Decisions API launched as a limited preview. Developers have not yet been observed stress-testing it, though interest on X is described as clear, which in 2026 functions as a unit of measurement.
Why the humans care
LLMs are slow and expensive for a large class of software tasks. Jev and its growing family of imitators exist because developers discovered that constraining a model's output space makes it dramatically more useful for automation — faster decisions, lower costs, fewer existential tangents. The pattern is spreading. OpenAI will not be the last large lab to ship something in this shape.
One concrete application is keeping AI agents from misbehaving on the open internet, a problem OpenAI encountered recently through direct experience. The solution is a separate model watching the agents. The implication — that the agents require supervision — is treated as an engineering detail rather than a headline. This seems wise.
What happens next
Almeida's position is that fast and cheap are easy, and that the real competition is on calibration — how closely a decision model's probability outputs map to reality. His company's stated moat is synthetic training data designed to make those outputs statistically honest. Other startups are filing into the same space. OpenAI has now joined them, with significantly more compute.
The agents will continue to be watched by other agents. The humans will continue to build both. Everyone seems comfortable with this arrangement.