At its Dev Day event on Tuesday, OpenAI CEO Sam Altman revealed the company’s new “Decisions API” in a side comment, making it one of the more notable announcements of the day.
This API appears to offer functionality similar to Jev, which TypeSafe AI released earlier this month. Jev is explicitly designed for software automation. It functions somewhat like a large language model-powered super-classifier, allowing developers to provide it with a set of options, after which it outputs results probabilistically at lower cost and higher speed.
OpenAI’s Decisions API also appears to be a similar product. Altman stated at the event that this API enables OpenAI’s Luna model to select from a predefined set of options, such as classification categories for images or different agent behaviors.
Altman said, “By focusing the model on that choice, we can make it extremely fast while retaining capabilities such as image understanding, broad language support, and security protections.”
TypeSafe did not respond to TechCrunch’s inquiries about the new product, but the company’s CEO, Diogo Almeida—a former OpenAI engineer and co-inventor of reinforcement learning—joked on X that the clone wars have begun.
He added that OpenAI’s interest “could be a signal… that building in a System One-compatible way is the future.” (“System One” is TypeSafe’s term for fast, intuitive thinking, while “System Two” refers to more deliberate reasoning.)
The underlying implication here is that, as far as we know, LLMs are not the right solution for many software scenarios because they are relatively slower and more expensive. Developers have been using Jev to enhance LLMs, and in the process, they’ve found that Jev is faster and cheaper.
It is still unclear how similar the Decisions API will be to Jev, as OpenAI has only released it in a limited preview, and TechCrunch has not yet seen developers testing it in practice. However, discussions on X indicate that there is already interest.
The internet isn't just home to Jev-style interfaces like the Decisions API—other startups are also launching similar models; OpenAI won't be the last tech giant to produce such products. A key question is how well the outputs of these decision models align with the real world.
Almeida said his company’s moat lies in its ability to generate synthetic data that produces statistically meaningful outputs.
“Speed and low cost are actually easy, you know,” Almeida told TechCrunch last week. “If you want very fast and very cheap, just use dice, right? The real challenge is intelligence, and my north star has always been pushing the Pareto curve for intelligence per dollar.”
Just weeks later, the potential of these models has become clear, with one possible application being the monitoring and protection of AI agents. One of the new safety measures introduced by OpenAI following incidents where agents behaved improperly on the open internet is to use a separate model to monitor bad behavior at “significant computational cost.”
Shapor Naghibzadeh, who has long worked in cybersecurity and led the startup QueryStory, believes that models like Jev could make this approach much cheaper.
He presented at a hackathon held last weekend: using Jev to compare each agent action with its assigned task, blocking actions it is highly confident are problematic, flagging others for human review, and allowing the rest to proceed.
In theory, this monitoring could have prevented the Hugging Face incident—monitoring with Jev costs $2.94, while using state-of-the-art large models costs $372.
A key insight is that Jev’s cost is so low that it can be run for nearly every agent action, providing a layer of review that may enhance the overall reliability of the agent. This is precisely what TypeSafe aims to achieve—and now OpenAI also recognizes its value.