Key facts
- OpenAI announced a new Decisions API at its Dev Day event.
- The Decisions API is designed to make AI models faster and cheaper by focusing them on predefined choices.
- TypeSafe AI's Jev model, released earlier this month, offers similar functionality for software automation.
- OpenAI's Luna model can use the Decisions API to classify images or choose agent behaviors.
- TypeSafe CEO Diogo Almeida suggested that building in a "System One compatible way" is the future.
- A model like Jev could monitor AI agent actions for security at a significantly lower cost than frontier LLMs.
OpenAI has introduced a new "Decisions API" that aims to make AI models faster and more cost-effective, drawing parallels to a recently released model from TypeSafe AI called Jev. Revealed by CEO Sam Altman at OpenAI's Dev Day event, the Decisions API is described as a way to provide AI models, such as OpenAI's Luna, with a predefined set of options to choose from, enabling rapid classification or decision-making.
This approach contrasts with the slower and more expensive nature of traditional large language models (LLMs) for many software automation tasks. TypeSafe AI's Jev model, built on an LLM, functions as a high-speed classifier that outputs probabilities for given choices. Developers have reportedly used Jev to augment LLMs, achieving faster and cheaper results.
Diogo Almeida, CEO of TypeSafe AI and a former OpenAI engineer, noted the similarity between the products and suggested that this focus on fast, intuitive thinking, which TypeSafe terms "System One," represents the future of AI development. While OpenAI has released its Decisions API as a limited preview, its existence signals a growing trend of specialized AI models designed for efficiency.
Beyond general software automation, a key application for these decision-focused models is in monitoring and securing AI agents. Following incidents of AI agents misbehaving online, OpenAI is exploring new security measures. Cybersecurity professionals like Shapor Naghibzadeh believe that models like Jev could provide this oversight much more cheaply. Naghibzadeh demonstrated a system using Jev to review agent actions, blocking potentially harmful ones and flagging others, at a fraction of the cost of using a frontier LLM.