OpenAI's Decisions API targets fast, cheap agent control
Original: OpenAI’s Jev clone could help the frontier lab stop its swarming agents
Why This Matters
Constrained-output classifiers are emerging as a distinct infrastructure layer for AI agent orchestration.
At Dev Day, OpenAI CEO Sam Altman unveiled a 'Decisions API' that gives its Luna model a predefined set of options to choose from — a product that closely mirrors Jev, a classifier-style model released by TypeSafe AI earlier in September 2026 for software automation tasks.
OpenAI's Dev Day brought several announcements, but a notable aside came when Altman described the new Decisions API: feed the model a constrained set of choices — image categories, agent behaviors — and it returns fast, cheap probability outputs. That's essentially what TypeSafe AI's Jev already does. Jev, built by former OpenAI engineer and RL co-inventor Diogo Almeida, functions as an LLM-powered super-classifier optimized for speed and low cost. Developers have been pairing it with conventional LLMs to cut latency and expense. Almeida responded on X with a joke about 'the clone wars,' while noting OpenAI's move may signal that 'building in a System One compatible way is the future' — TypeSafe's shorthand for fast, intuitive processing as opposed to slow deliberate reasoning. The Decisions API is currently in limited preview, so direct comparisons are premature. Almeida argues TypeSafe's real advantage is calibration: the synthetic data it generates to produce statistically grounded outputs. 'Fast and cheap is very easy,' he told TechCrunch. 'Intelligence is the hard part.' OpenAI isn't alone — other startups are shipping similar models, and the underlying use case (controlling and securing swarms of AI agents) is becoming a legitimate product category.