At OpenAI’s Dev Day on Tuesday, CEO Sam Altman briefly introduced the company’s new Decisions API. According to his remarks, the API provides functionality similar to TypeSafe AI’s Jev model, released earlier this month: a lightweight decision layer built on an LLM that delivers fast, inexpensive probabilistic choices.
What the Decisions API does
Altman said the Decisions API lets developers provide the lab’s Luna model with a predefined set of options to choose between — for example, categories for classifying an image or alternative agent behaviors. By narrowing the model’s task to a specific choice, OpenAI says it can run the model extremely quickly while retaining capabilities such as image understanding, broad language support, and safety protections.
Altman: “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.”
Relation to Jev and industry reaction
The Decisions API resembles TypeSafe AI’s Jev, which functions as a sort of super-powered classifier: developers feed Jev a set of choices and receive probabilities cheaply and at high speed. TypeSafe did not respond to TechCrunch’s questions about OpenAI’s new product, but TypeSafe CEO Diogo Almeida — described in the report as a former OpenAI engineer and a co-inventor of reinforcement learning — joked on X that this marks the start of the “clone wars.” Almeida also suggested OpenAI’s interest could indicate that building in a “System One” (TypeSafe’s term for fast, intuitive thinking) compatible way may be the future.
The underlying implication is that traditional LLMs are often too slow and expensive for many software tasks. Developers using Jev alongside LLMs have reported it is faster and cheaper. OpenAI released Decisions API as a limited preview, so its detailed behavior compared with Jev is not yet clear, but conversations on X show notable interest.
Application to agent monitoring and safety
One prominent use case for these decision models is monitoring and securing autonomous agents. After a series of incidents where OpenAI agents misbehaved online, the company added security measures that include using a separate model to watch for bad actions — an approach that OpenAI describes as incurring “significant compute cost.”
Shapor Naghibzadeh, a long-time cybersecurity professional who leads the startup QueryStory, believes a Jev-like model can perform such monitoring much more cheaply. In a demo he built for a hackathon last weekend, he used Jev to check each agentic action against its assigned task: actions with high confidence of being bad were blocked, others were flagged for review, and the rest were allowed.
In that demo, such monitoring could theoretically have prevented the Hugging Face incident — and the cost difference was stark: the Jev-based check cost $2.94, versus $372 with a frontier LLM.
Outlook and key questions
Multiple startups are rolling out Jev-like decision APIs, and OpenAI will likely not be the last major company to enter this space. A central question going forward is how well calibrated these decision models’ probabilistic outputs are for real-world decisions — in other words, how reliable and accurate their probability estimates will be.
According to Almeida, TypeSafe’s moat lies in the synthetic data it generates to produce statistically useful outputs. He told TechCrunch: “Fast and cheap is very easy, you know. 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.”
In the near term, these fast, inexpensive decision models appear promising, particularly for agent oversight where they may be cheap enough to run on every agentic action and add an additional layer of review to improve overall reliability.
