One of the most interesting announcements at OpenAI’s Dev Day event on Tuesday was CEO Sam Altman’s aside revealing the company’s new “Decisions API.”
This API apparently provides similar functionality to Jev, a model explicitly designed for software automation, released by TypeSafe AI earlier this month. It’s a kind of super-powerful classifier built on top of LLM that allows developers to give Jev a set of choices and output them as probabilities cheaply and quickly.
OpenAI’s Decisions API appears to be a similar product. At the event, Altman described the API as a way to provide the lab’s Luna model with a set of predefined options from which to choose, such as images and categories to categorize different agent behaviors.
“By focusing the model on that selection, we can make it very fast while still maintaining features like image understanding, extensive language support, and safety protections,” Altman said.
TypeSafe did not respond to TechCrunch’s questions about the new product, but CEO Diogo Almeida, a former OpenAI engineer who co-invented reinforcement learning, joked about the start of the Clone Wars with X.
He added that the interest in OpenAI could be “a sign that building in a System One compatible way is the future.” (“System 1” is TypeSafe terminology for fast, intuitive thinking, and “System 2” applies to deliberate reasoning.)
The implication here is that LLM as we know it is relatively slow and expensive, so it is not a suitable solution for many software. Developers have been using Jev to power LLM, and in doing so have found that LLM is faster and cheaper.
It is unclear how similar the Decisions API is to Jev. Released by OpenAI in limited preview, TechCrunch has so far not found any developers putting the Decisions API through its paces. However, based on the conversation at X, there is clearly interest.
The Decisions API is not the only Jev-like API on the Internet. Other startups are rolling out similar models. OpenAI is not the last tech giant to develop this. A key question is how well the output of each of these decision-making models is calibrated to real life.
Almeida says his company’s moats are synthetic data created to produce statistically useful output.
“It’s a no-brainer that it’s fast and cheap,” Almeida told TechCrunch last week. “If you want to do it really fast and cheap, why not just use dice? Intelligence is the hard part, and my north star is always pushing up the Pareto curve of intelligence per dollar.”
After just a few weeks, it seems clear that these models have a future. One of its applications is the monitoring and protection of AI agents. In response to a series of incidents in which agents misbehaved on the open internet, one of the new security measures OpenAI has introduced is to use a different model to monitor for fraud at “significant computational cost.”
Shapor Naghibzadeh, a longtime cybersecurity expert who leads the startup QueryStory, believes models like Jev can do that much more cheaply.
He created a demo for a hackathon held last weekend. In this demo, we use Jev to check each agent’s actions against a given task, block the actions we believe are bad, flag others for review, and allow the rest.
In theory, such surveillance could have prevented the “hugface” incident. That kind of monitoring costs $2.94 at Jev, compared to $372 at Frontier LLM.
The key observation is that Jev is probably cheap enough to run on all agent actions and provides a review layer that can significantly improve agent reliability. This is the kind of thing TypeSafe wanted to achieve, and OpenAI recognized its value as well.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
