OpenAI has introduced a new tool aimed at improving the efficiency and safety of autonomous AI agents, signaling a potential shift in how developers manage large-scale agent swarms. During its Dev Day event on Tuesday, CEO Sam Altman revealed the "Decisions API," a product that appears to offer functionality similar to Jev, a model recently released by startup TypeSafe AI.
What Happened
The Decisions API is designed to provide a predefined set of options for OpenAI’s Luna model to choose between, such as image categories or specific agent behaviors. Altman described the API as a way to focus the model on specific choices, thereby making it "extremely fast while keeping capabilities like image understanding, broad language support, and safety protections." The API is currently available as a limited preview.
The tool draws immediate comparisons to Jev, a model released by TypeSafe AI earlier this month. Jev is explicitly designed for software automation, functioning as a high-speed, low-cost classifier built on an LLM that outputs probabilities for a given set of choices. While TypeSafe did not respond to direct questions about the new competitor, CEO Diogo Almeida, a former OpenAI engineer, commented on social media that OpenAI’s move could indicate that "building in a System One compatible way is the future." TypeSafe uses the term "System One" to describe fast, intuitive thinking, as opposed to "System 2," which refers to deliberate reasoning.
Why It Matters
The emergence of tools like Jev and the Decisions API addresses a critical bottleneck in current AI agent development: the high latency and cost of using frontier LLMs for routine decision-making. Developers have increasingly turned to specialized, faster models to augment LLMs, finding that these alternatives are significantly cheaper and quicker for classification tasks. A key application for these tools is monitoring and securing AI agents, particularly as incidents of agents misbehaving on the open internet have prompted labs to implement costly oversight measures.
Shapor Naghibzadeh, founder of cybersecurity startup QueryStory, demonstrated how a model like Jev could be used to check each agentic action against its assigned task. In his hackathon demo, the model blocked actions with high confidence of being incorrect, flagged others for review, and permitted the rest. According to the source, this type of monitoring cost $2.94 using Jev, compared to $372 using a frontier LLM. This price differential suggests that specialized decision models could enable comprehensive, real-time safety checks on every agentic action, potentially preventing incidents like the recent Hugging Face issue.
The Bottom Line
While it remains to be seen how closely OpenAI’s Decisions API will match Jev’s performance, the move underscores a growing industry consensus that large, general-purpose models are inefficient for simple classification and control tasks. As other startups and tech giants roll out similar "System One" style models, the focus is shifting toward balancing speed, cost, and calibration. Diogo Almeida noted that while making models fast and cheap is straightforward, the challenge lies in maintaining intelligence, stating his goal is to push the "intelligence-per-dollar Pareto curve."