At the Dev Day event held on Tuesday, CEO Sam Altman disclosed the company's new “Decisions API” in a side note, which became one of the more notable announcements of the day.

This API seems to offer similar functionality to the Jev that was released earlier this month by TypeSafe AI. Jev is explicitly designed for software automation. It's somewhat like a super classifier driven by large language models; developers can provide it with a set of options, and it will output results in the form of probabilities, at a lower cost and higher speed.

The Decisions API of OpenAI also seems to be a similar product. Altman stated during the event that this API allows the company's Luna model to make choices from a pre-set set of options, such as categories for image classification, or different proxy behaviors.

Altman said, "By focusing the model on that selection, we can make it extremely fast while still retaining capabilities such as image understanding, broad language support, and security protection."

TypeSafe did not respond to TechCrunch's questions about this 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 war has begun.

He added that the interest in OpenAI "may be a signal... indicating that building in a way compatible with System One is the future." ("System One" is a term used by TypeSafe for fast, intuitive thinking, while "System 2" is used to refer to more cautious reasoning.)

The subtext here is that, as far as we know, LLM is not the right solution for many software scenarios, as it is relatively slower and more expensive. Developers have been using Jev to enhance LLM, and in the process, they found that it is faster and cheaper.

It is still unclear how similar Decisions API will be to Jev, as OpenAI has only been released in a limited preview version, and so far, TechCrunch has not seen any actual testing by developers. However, based on the discussions on X, there is clearly already some interest.

In the internet, there's not just the Decisions API type of Jev interface; other startups are also launching similar models; and OpenAI will certainly not be the last tech giant to produce such products. A key question is, to what extent do the outputs of these decision-making models match the real world?

Almeida said that the moat of his company lies in its ability to generate synthetic data, which can produce statistically useful outputs.

“It’s actually easy to be fast and cheap, you know,” Almeida told TechCrunch last week. “If you want it to be extremely fast and extremely cheap, then use dice, right? The real challenge lies in intelligence, and my North Star has always been to optimize the Pareto curve for every dollar of intelligence spent.”

Just a few weeks later, the prospects for these models seemed quite clear, and one possible application was in monitoring and protecting AI proxies. One of the new security measures introduced by OpenAI after a series of incidents where proxies behaved improperly on the open internet was to use a separate model to monitor for maladaptive behavior at a “significant computational cost”.

Shapor Naghibzadeh, who has been engaged in cybersecurity for a long time and leads the startup QueryStory, believes that models like Jev can make this approach much more cost-effective.

He prepared a demonstration for a hackathon held last weekend: using Jev to compare each proxy action with the task it was assigned to, blocking those actions he was highly certain were problematic, marking the others for manual review, and allowing the rest to proceed.

Theoretically, this kind of monitoring could have prevented the Hugging Face incident – and the cost of using Jev for such monitoring is $2.94, whereas using cutting-edge large models would cost $372.

A key observation is that the cost of Jev is so low that it can be executed almost for every proxy action, which provides an extra layer of oversight and may enhance the overall reliability of the proxy. This is precisely the goal that TypeSafe aims to achieve, and now OpenAI also sees the value in this approach.