OpenAI Launches Decisions API for Faster, Cheaper AI Choices as Agent Monitoring Costs Drop to $2.94
Updated
Updated · TechCrunch · Sep 30
OpenAI Launches Decisions API for Faster, Cheaper AI Choices as Agent Monitoring Costs Drop to $2.94
3 articles · Updated · TechCrunch · Sep 30
Summary
OpenAI unveiled a limited-preview Decisions API at Dev Day, letting its Luna model choose among predefined options for tasks like image classification and agent behavior selection.
Sam Altman said constraining the model to fixed choices makes it extremely fast while preserving image understanding, broad language support and safety protections—targeting a key weakness of slower, costlier general LLMs.
The product resembles TypeSafe AI’s Jev, released earlier this month for software automation; Jev users have been using that approach to augment LLMs with faster, cheaper probability-based decisions.
One likely use is supervising AI agents: QueryStory’s Shapor Naghibzadeh built a Jev-based demo that checked every agent action against its task, with monitoring estimated at $2.94 versus $372 for a frontier LLM.
That cost gap suggests decision models could become a standard reliability layer for agentic systems, though their real-world value will hinge on how well their outputs are calibrated.
Is OpenAI’s new Decisions API a real breakthrough for cheaper, safer agents—or just a fast-follow move into a market Jev already defined?
Are fast “System One” decision models becoming the real backbone of AI products, replacing expensive frontier models for most everyday choices?
Can a low-cost decision layer really stop dangerous AI agent actions, or does true safety still depend on code-level permissions and external watchdogs?