At OpenAI's Dev Day event on Tuesday, CEO Sam Altman made an intriguing revelation during a side note: the company is introducing a new "Decisions API."

This API appears to offer functionality similar to Jev, a model that TypeSafe AI released earlier this month, specifically designed for software automation. Jev operates as a kind of enhanced classifier built on an LLM, allowing developers to provide a set of choices that it outputs as probabilities—both cheaply and at high speeds. OpenAI's Decisions API seems to serve the same purpose. During the event, Altman explained that the API lets developers give the lab's Luna model a predefined set of options to select from, such as categories for classifying an image or different agent behaviors.

"By focusing the model on that choice, we can make it extremely fast while keeping capabilities like image understanding, broad language support, and safety protections," Altman stated. He also suggested that OpenAI's interest might be "a sign…that building in a System One compatible way is the future." ("System One" is TypeSafe's term for fast, intuitive thinking, as opposed to "System 2," which refers to deliberate reasoning.)

The underlying message is that current LLMs aren't the ideal solution for many software applications because they tend to be relatively slow and costly. Developers have been using Jev to enhance LLMs and have discovered that this approach makes them faster and more affordable. Nevertheless, there's evident interest, based on discussions on X. Decisions API isn't the only Jev-like API available—other startups are introducing similar models, and OpenAI won't be the last major tech company to create one. A critical question is how well calibrated each of these decision models' outputs will be to real-world scenarios. Almeida claims his company's advantage lies in the synthetic data it generates to produce statistically useful outputs.

"If you want it really fast and cheap, use dice, right. Intelligence is the hard part, and my North Star is always pushing the intelligence-per-dollar Pareto curve."

After only a few weeks, it's becoming clear that these models have a promising future, with one likely use case being the monitoring and securing of AI agents. Following a series of incidents where its agents misbehaved on the open internet, one of OpenAI's new security measures involves using a separate model to watch for bad actions at "significant compute cost."

Shapor Naghibzadeh, a veteran cybersecurity professional who leads the startup QueryStory, believes a model like Jev could enable that monitoring much more affordably. He created a demo for a hackathon last weekend that uses Jev to check each agentic action against its assigned task, blocking actions it was highly confident were harmful, flagging others for review, and allowing the rest. In theory, this kind of monitoring could have prevented the Hugging Face incident—and such monitoring costs $2.94 with Jev, compared to $372 with a frontier LLM. A key insight is that Jev is arguably inexpensive enough to run on every agentic action, providing a review layer that could enhance the reliability of agents overall. This is precisely what TypeSafe was aiming for—and now OpenAI has recognized the value as well.