At OpenAI’s recent Dev Day event, CEO Sam Altman introduced a new product called the “Decisions API,” which functions similarly to a model named Jev, released earlier this month by TypeSafe AI. Jev, designed for software automation, leverages large language models (LLMs) to quickly generate probability-based choices across options, offering a faster and more cost-effective alternative to typical LLM operations. OpenAI’s Decisions API aims to replicate this approach by focusing on speedy decision-making within predefined option sets, maintaining robust capabilities like image analysis and language understanding while improving speed and affordability.

The emergence of OpenAI’s Decisions API highlights a growing recognition that traditional LLMs, while powerful, can be too slow and expensive for various software tasks. TypeSafe’s CEO, Diogo Almeida, who is also a former OpenAI engineer, acknowledged this trend, noting that fast, cheap intelligence models like Jev exemplify the future direction where intuitive “System One” thinking is favored over slower, deliberate processing. This shift has attracted attention in developer communities, though OpenAI’s version remains in limited preview with few public demonstrations available so far.

One key potential use case for these fast decision models is enhancing AI security by monitoring agent behaviors. After incidents involving misbehaving AI agents, OpenAI has employed additional models to oversee actions, but the high computational costs have been a barrier. Startups like QueryStory are using Jev to run efficient, continuous checks on agent decisions, blocking or flagging inappropriate actions with significantly reduced expense—$2.94 compared to $372 with conventional models. This approach could help prevent issues like unauthorized data exposure, bolstering the reliability and safety of autonomous AI systems.

The development of Jev-like models by both startups and major companies indicates a competitive and rapidly evolving landscape in AI automation technology. The core challenge remains balancing speed, cost, and intelligence quality, with TypeSafe claiming an edge due to its use of synthetic data for output calibration. As these models gain traction, they promise to accelerate automation processes while providing a practical means to monitor and control multi-agent AI deployments, a crucial step for scaling advanced AI applications securely.