In a significant development for the artificial intelligence landscape, OpenAI has introduced its new “Decisions API,” a system that bears a striking resemblance to TypeSafe AI's innovative Jev model. This announcement, made during OpenAI's recent Dev Day, underscores the growing demand for efficient and economical AI solutions in software automation and intelligent agent oversight.
OpenAI Introduces Decisions API, Echoing TypeSafe AI's Jev for Enhanced AI Control
During OpenAI’s annual developer conference on Tuesday, September 30, 2026, CEO Sam Altman briefly highlighted a new offering: the “Decisions API.” This API’s functionality appears strikingly similar to Jev, a cutting-edge model launched earlier this month by TypeSafe AI, specifically engineered to streamline software automation processes. Jev, functioning as a sophisticated classifier powered by large language models (LLMs), empowers developers to present a range of options, from which it rapidly and economically calculates probabilities.
OpenAI’s Decisions API reportedly offers an analogous capability. Altman explained that the API allows their Luna model to select from a predefined set of choices, whether classifying an image or dictating agent behaviors. He emphasized that by narrowing the model's focus to such choices, OpenAI can achieve exceptional speed while maintaining critical attributes like image comprehension, comprehensive language support, and robust safety protocols. While TypeSafe AI declined to comment on this competitive development, CEO Diogo Almeida, a former OpenAI engineer and co-creator of reinforcement learning, playfully remarked on social media about the onset of "clone wars," suggesting that OpenAI's interest validates the future of "System One compatible" AI – TypeSafe's term for intuitive, rapid thinking, contrasting with the more deliberate "System 2" reasoning.
The underlying motivation for such models stems from the inherent limitations of conventional LLMs, which are often slow and resource-intensive for many software applications. Developers have leveraged Jev to augment LLMs, experiencing notable improvements in speed and cost efficiency. The exact parallels between Decisions API and Jev remain to be fully evaluated, as OpenAI's offering is currently in limited preview. However, preliminary discussions on platforms like X indicate a strong interest within the developer community.
The Decisions API is not an isolated innovation; several startups are also developing similar decision-making models. A pivotal challenge for these systems is ensuring the accuracy and real-world applicability of their outputs. Almeida states that TypeSafe's competitive edge lies in its ability to generate synthetic data, which is crucial for producing statistically reliable results. He previously articulated that while achieving speed and affordability is straightforward, true intelligence—optimizing the intelligence-per-dollar ratio—is the more complex endeavor.
Looking ahead, these rapid decision models are poised to revolutionize AI agent monitoring and security. One immediate application is in overseeing AI agents, particularly in light of recent incidents where AI agents exhibited undesirable behaviors online. Shapor Naghibzadeh, a cybersecurity veteran and leader of QueryStory, demonstrated a proof-of-concept using Jev at a recent hackathon. This demo illustrated how Jev could verify each agentic action against its assigned task, automatically blocking high-risk actions, flagging others for human review, and permitting safe operations. This kind of monitoring could have averted past incidents, and critically, it offers a stark cost advantage: monitoring an agent costs an estimated $2.94 with Jev, compared to $372 with a frontier LLM. The ability to cheaply run real-time checks on every agent action promises a new era of reliability and safety for AI agents, a vision that both TypeSafe and now OpenAI are actively pursuing.
The advent of sophisticated decision-making APIs like OpenAI's and TypeSafe AI's marks a significant leap in the practical application of artificial intelligence. It highlights a critical shift towards more efficient, cost-effective, and secure AI systems. As AI agents become increasingly autonomous, the ability to monitor and control their actions in real-time, at a low cost, is paramount. This innovation not only promises to enhance the reliability of AI but also opens new avenues for addressing complex challenges in fields ranging from cybersecurity to operational automation. The collaborative yet competitive landscape among AI developers suggests a rapid evolution in how we interact with and manage intelligent systems, ultimately aiming for a future where AI operates with greater precision and trustworthiness.
