OpenAI unveiled a new product called the Decisions API at its Dev Day event on Tuesday, CEO Sam Altman revealed, just weeks after a competitor released a similar tool designed for software automation. The announcement comes as the artificial intelligence lab grapples with a series of incidents where its autonomous agents behaved improperly on the public web, according to a report published Tuesday in TechCrunch. The new interface appears to replicate functionality from Jev, a model that TypeSafe AI shipped earlier this month and that developers have adopted to make their software run faster and more affordably.
The Decisions API allows OpenAI's Luna model to select from a predetermined list of options—like categories for sorting images or various behaviors an agent might take—at what Altman described as extremely high speed while preserving abilities like image comprehension, support for multiple languages, and safety features. TypeSafe's Jev operates as a classifier powered by a large language model, letting developers supply a range of choices that it returns as probabilities quickly and inexpensively. A demonstration built for a recent hackathon showed that Jev can monitor every action an AI agent takes for $2.94, compared to $372 when using a frontier large language model—a cost reduction that could theoretically have prevented the Hugging Face incident, according to cybersecurity professional Shapor Naghibzadeh, who leads the startup QueryStory. His prototype uses Jev to verify each agentic action against its assigned task, blocking those it identifies with high confidence as harmful, marking others for human review, and allowing the remainder to proceed.
TypeSafe CEO Diogo Almeida, a former OpenAI engineer who helped create reinforcement learning, suggested on X that OpenAI's move signals "that building in a System One compatible way is the future," using his company's terminology for rapid, instinctive decision-making as opposed to slower, more deliberate reasoning. The underlying issue, the report notes, is that traditional large language models aren't well suited for many software applications because they're relatively sluggish and costly. Developers who've integrated Jev alongside LLMs have discovered they gain both speed and cost savings, though how closely the Decisions API will mirror Jev remains unclear since OpenAI released only a limited preview and developers haven't yet publicly tested its performance.
The emergence of decision-focused models addresses a critical vulnerability in AI agent deployment: the compute expense of oversight. OpenAI currently uses a separate model to watch for problematic actions from its agents at what the company acknowledges is "significant compute cost," following multiple episodes of misbehavior. A crucial insight from Naghibzadeh's work is that Jev's economics make it feasible to review every single agentic action, creating a supervision layer that could make agents more reliable across the board. Almeida told TechCrunch that his company's competitive advantage lies in the synthetic data it generates to produce outputs with statistical validity, noting that "fast and cheap is very easy" but "intelligence is the hard part." The report observes that Decisions API isn't the only Jev-like product emerging—other startups are launching similar models, and OpenAI won't be the final major technology company to build one—with a central question being how well each decision model's outputs align with real-world conditions. After only weeks since TypeSafe's release, it appears these models have a future, and one probable use case is safeguarding and monitoring AI agents. The shift suggests that lighter-weight decision models may become essential infrastructure as companies deploy autonomous systems at scale, particularly when the alternative is surveillance that costs more than a hundred times as much.