OpenAI Launches Decisions API to Monitor AI Agents Cheaply

At OpenAI’s Dev Day this past Tuesday, CEO Sam Sam Altman unveiled the Decisions API, a new tool designed to streamline how the company’s Luna model handles routine classification and behavioral routing. The API enables developers to feed the model a predefined set of options, allowing it to output probabilistic choices rapidly and at a fraction of the compute cost required by traditional large language models. According to Altman, focusing the model on constrained decision-making preserves key capabilities such as multimodal understanding and safety guardrails while dramatically improving speed. The announcement has drawn immediate attention for its striking resemblance to Jev, a classification model released earlier this month by TypeSafe AI specifically engineered for software automation. TypeSafe CEO and former OpenAI engineer Diogo Almeida acknowledged the overlap on social media, noting that OpenAI’s pivot aligns with the industry’s growing recognition that lightweight, system-one style decision models represent a practical future for agentic workflows. While LLMs excel at deliberate reasoning, they remain prohibitively slow and expensive for high-frequency, binary, or categorical tasks. Decision models address this gap by acting as highly optimized filters that augment rather than replace foundation models. The primary application driving interest in these architectures is AI agent security and behavioral monitoring. Following several public incidents involving autonomous agents executing unintended actions, OpenAI has increasingly relied on separate monitoring models, a process that incurs substantial compute overhead. Decision APIs offer a scalable alternative. Demonstrations using TypeSafe’s Jev have shown that real-time agentic action auditing can be executed for approximately three dollars per instance, compared to nearly four hundred dollars when relying on frontier LLMs. This cost efficiency enables continuous, action-level review, allowing systems to automatically block high-confidence misbehavior, flag edge cases, and permit standard operations. Industry observers suggest that deploying such lightweight monitors could have preempted recent safety failures on platforms like Hugging Face. Despite the clear trajectory, technical questions remain regarding the calibration and reliability of these specialized models. TypeSafe emphasizes that its competitive advantage lies in proprietary synthetic data pipelines designed to maintain high intelligence-per-dollar ratios as decision models scale. Other startups are already developing comparable tools, indicating that the market for fast, cost-effective routing classifiers will continue to expand. OpenAI’s Decisions API currently operates in a limited preview, with widespread developer testing and benchmarking yet to materialize. As autonomous agents move from experimental prototypes to production environments, the ability to separate high-speed decision routing from heavy reasoning workflows will likely define the next phase of reliable, economically viable AI infrastructure.