OpenAI's Decisions API takes aim at TypeSafe's Jev for cheap agent control

OpenAI previews a Jev-style Decisions API

One of the more quietly significant announcements at OpenAI's Dev Day event on Tuesday came in an aside from CEO Sam Altman: a new Decisions API. According to TechCrunch, the product appears to offer functionality similar to Jev, a model released by startup TypeSafe AI earlier this month that is explicitly designed for software automation.

Jev is effectively a supercharged classifier built on an LLM. A developer supplies a set of choices, and the model returns them as probabilities, cheaply and at high speed. Altman described the Decisions API in the same terms, as a way to give OpenAI's Luna model a predefined set of options to pick between, 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 said at the event.

TypeSafe sees a clone war coming

TypeSafe did not respond to TechCrunch's questions about the new product, but its CEO, Diogo Almeida, a former OpenAI engineer, joked on X about the beginning of the clone wars. He also suggested OpenAI's interest could be a sign "that building in a System One compatible way is the future." System One is TypeSafe's term of art for fast, intuitive thinking, as opposed to System 2, which it applies to deliberate reasoning.

The subtext, as TechCrunch points out, is that general-purpose LLMs are the wrong tool for a lot of software because they are comparatively slow and expensive. Developers have been using Jev to augment LLMs and, in doing so, have found their systems run faster and cost less.

A cheap way to police agents

One likely application for this class of model is monitoring and securing AI agents. Following a series of incidents in which its agents misbehaved on the open internet, OpenAI has adopted new security measures that include a separate model watching for bad actions at "significant compute cost," TechCrunch reports.

Shapor Naghibzadeh, a long-time cybersecurity professional who leads the startup QueryStory, built a hackathon demo last weekend that uses Jev to check each agentic action against the task the agent was given: blocking actions it identifies with high confidence as bad, flagging others for review, and permitting the rest. In theory, TechCrunch notes, that kind of monitoring could have stopped the Hugging Face incident, and the report puts the cost at 2.94 dollars with Jev versus 372 dollars with a frontier LLM.

The key observation is that Jev is arguably cheap enough to run on every single action an agent takes, adding a layer of review that could improve the reliability of agents writ large.

Open questions

How closely the Decisions API will match Jev is unclear. OpenAI released it as a limited preview, and TechCrunch says it has not yet spotted developers putting it through its paces, though conversations on X suggest plenty of interest. Nor is OpenAI alone: other startups are rolling out similar models, and it likely will not be the last tech giant to build one.

A central question for all of these decision models, per TechCrunch, is calibration, meaning how well their probability outputs correspond to real life. Almeida argues that TypeSafe's moat is the synthetic data it creates to generate statistically useful outputs.

"Fast and cheap is very easy, you know," Almeida told TechCrunch. "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."

Why it matters

OpenAI adopting a small startup's idea within weeks is itself a signal: constrained, choice-based models look set to become a standard complement to general-purpose LLMs, handling the narrow, high-volume decisions that dominate agent workloads at a fraction of the usual cost and latency.

The economics matter most for safety. If checking every action an agent takes costs a few dollars instead of hundreds, oversight stops being a scarce resource rationed with spare compute and becomes a default layer beneath every agent, or swarm of agents, that a lab ships.

The competitive picture shifts too. With OpenAI and other giants copying the format, TypeSafe's defense rests not on being first but on calibration and what Almeida calls the intelligence-per-dollar curve, the difference between a fast model and a genuinely useful one.