OpenAI’s Jev Clone May Help Frontiers Lab Control Swarming Agents

Among the most notable announcements at OpenAI’s Dev Day on Tuesday was a brief remark from CEO Sam Altman regarding the launch of the company’s new “Decisions API.”

This API appears to offer capabilities similar to Jev, a model introduced earlier this month by TypeSafe AI specifically for software automation. Functioning as a high-speed, cost-effective classifier powered by an LLM, Jev allows developers to input a set of options and receive probability-based outputs rapidly.

OpenAI’s Decisions API seems to serve the same purpose. During the event, Altman explained that the API enables the lab’s Luna model to select from a predefined set of choices, such as image classification categories or specific agent behaviors.

“By narrowing the model’s focus to a specific choice, we can achieve extreme speed while maintaining capabilities like image understanding, broad language support, and safety protections,” Altman stated.

TypeSafe did not immediately respond to TechCrunch’s inquiries about the new product. However, CEO Diogo Almeida, a former OpenAI engineer and co-inventor of reinforcement learning, joked on X about the potential onset of a “clone war.”

He suggested that OpenAI’s interest might indicate that building in a “System One” compatible manner represents the future. (“System One” is TypeSafe’s terminology for fast, intuitive processing, as opposed to “System 2,” which refers to deliberate reasoning.)

The underlying implication is that current LLMs are not the optimal solution for many software tasks due to their relative slowness and high costs. Developers have been using Jev to augment LLMs, discovering that this approach yields faster and more cost-effective results.

It remains unclear how closely the Decisions API mirrors Jev, as OpenAI has only released it as a limited preview. Consequently, TechCrunch has not yet observed developers extensively testing it. Nevertheless, conversations on X suggest significant industry interest.

The Decisions API is not the only Jev-like offering available; other startups are launching similar models, and OpenAI is unlikely to be the last tech giant to do so. A critical question remains regarding how well these decision models’ outputs align with real-world scenarios.

Almeida argues that his company’s competitive advantage lies in the synthetic data it generates to produce statistically robust outputs.

“Achieving speed and low cost is straightforward,” Almeida told TechCrunch last week. “If you prioritize extreme speed and affordability, you can use dice. The real challenge is intelligence, and my primary goal is always to push the intelligence-per-dollar Pareto frontier.”

Within just weeks, it has become evident that these models have a promising future, with one likely application being the monitoring and security of AI agents. Following several incidents where OpenAI’s agents behaved unpredictably on the open internet, one of the company’s new security measures involves using a separate model to detect harmful actions, albeit at a “significant compute cost.”

Shapor Naghibzadeh, a cybersecurity veteran leading the startup QueryStory, believes that a model like Jev could make this process far more affordable.

He created a demo for a hackathon held last weekend that utilizes Jev to evaluate each agentic action against its assigned task. The system blocks actions it deems highly likely to be malicious, flags others for review, and permits the rest.

Theoretically, such monitoring could have prevented the Hugging Face incident. Moreover, monitoring via Jev costs $2.94, compared to $372 when using a frontier LLM.

A key insight is that Jev is inexpensive enough to run on every agentic action, providing a layer of review that could enhance the overall reliability of agents. This aligns with TypeSafe’s original vision—and now OpenAI has recognized its value as well.