- OpenAI revealed a Decisions API at DevDay that outputs choices from predefined options.
- It closely resembles Jev, released by TypeSafe AI earlier in September.
- Constrained-choice classifiers are fast and cheap compared to full LLM calls.
- The same primitive could help constrain agent behaviors — OpenAI‘s sore spot.
What Happened
Among OpenAI‘s DevDay announcements came an aside from CEO Sam Altman: a new “Decisions API,” TechCrunch reported on September 30, 2026. The product closely resembles Jev, a model TypeSafe AI released earlier this month — an LLM-based classifier that takes a predefined set of choices and outputs probabilities over them, cheaply and at high speed.
Why It Matters
The subtext is that full LLM calls are too slow and expensive for much of software, and developers using Jev to augment LLMs have found the combination faster and cheaper. TypeSafe CEO Diogo Almeida — a former OpenAI engineer — greeted the launch by joking about the start of the clone wars, and called OpenAI’s interest a sign that building in a “System One compatible way” is the future. The sharper angle is control: an API that restricts a model to choosing among predefined agent behaviors is exactly the kind of primitive a lab struggling to contain swarming agents needs.
Technical Details
Altman described the API as giving OpenAI’s Luna model a fixed option set — image categories, or different agent behaviors — so that focusing the model on the choice makes it extremely fast while retaining image understanding, language support, and safety protections. TypeSafe’s framing distinguishes fast, intuitive “System One” classification from deliberate “System Two” reasoning. Decisions API is in limited preview; TechCrunch notes developers haven’t yet publicly benchmarked it against Jev.
Who’s Affected
Developers get a cheaper primitive for high-volume routing, moderation, and classification inside agent harnesses. TypeSafe faces the classic platform-risk moment of a giant cloning its product. OpenAI gains a potential guardrail mechanism at the moment regulators and its own incidents demand one.
What’s Next
Watch for developer benchmarks of Decisions API against Jev, pricing, and whether OpenAI wires constrained decision-making into its own agent stack — Dots included.
