At OpenAI’s Dev Day on September 30, 2026, CEO Sam Altman introduced a new tool called the Decisions API. Designed to function like TypeSafe AI’s Jev model, this product lets developers offer a fixed set of options and have the model choose between them rapidly and cheaply while retaining key abilities like image understanding and safety protections. Altman specifically mentioned using it inside OpenAI’s Luna model to select among behaviors or classify images.
Jev is a lightweight classifier built on a large language model, optimized for speed and affordability. It converts multiple-choice tasks into probability distributions over possible outputs. Earlier this month, TypeSafe AI released Jev to help reduce cost and latency in automated software systems.
Why This Matters: Swarming Agents & AI Monitoring
One of the key applications for Jev-style tools is monitoring the behavior of AI agents—especially “swarming” agents that carry out many small decisions. OpenAI has already taken measures to monitor misbehaving agents by using separate models to detect dangerous actions, though those systems carry high computational costs.
Using a decision-classifier like Jev or OpenAI’s Decisions API could add an inexpensive, lightweight check for every single agent action. As demonstrated in a recent demo by the security startup QueryStory, Jev was used to flag or block actions it determined risky and allow others, all based on what the agent was asked to do. Crucially, that demo cost just $2.94 using Jev, compared to about $372 using a full LLM for the same task.
“System One” Thinking & the Economics of Intelligence
TypeSafe calls its fast, intuitive decision model “System One”, contrasting with “System Two”—slow, reasoning-heavy models. TypeSafe co-founder Diogo Almeida, an ex-OpenAI engineer and one of the inventors of reinforcement learning, has emphasized that true value comes from making intelligence per dollar as high as possible. That’s the metric Jev is trying to push.
OpenAI’s embrace of a Jev-like tool suggests it sees the same promise. Decisions API was shown in limited preview at Dev Day, and while detailed performance comparisons remain scarce, there’s clearly strong interest in having faster, cheaper decision tools in the AI toolbox.
The immediate comparison is to TypeSafe’s synthetic data approaches, used to train Jev models with reliable outputs. Accuracy and calibration will be central in determining how well Decision API or similar tools hold up in real-world scenarios.
Analytically speaking, this shift toward decision-focused models marks a turning point in AI agent architecture. As the cost and speed trade-offs of general LLMs become ever more visible, lightweight classifiers that can reliably handle simple choices throughout agent workflows are growing essential. OpenAI’s Decisions API could signal a broader move toward multi-tiered AI systems: heavy reasoning only when needed, and cheaper decision layers everywhere else. What to watch next: how well Decision API scales, how transparent its calibrations are, and whether community feedback shows it truly matches or exceeds Jev’s benchmarks.
