OpenAI's Decisions API: betting on agents that decide more cheaply, not necessarily more smartly

By Zendoric · September 30, 2026.

OpenAI unveiled on Tuesday, during its Dev Day, a limited preview of the Decisions API: a tool that lets its models — in this case the one called Luna — choose from a closed set of options that the developer defines in advance, rather than generating an open-ended response. Sam Altman presented it as a route for tasks such as classifying images or deciding what behavior an AI agent should adopt, with the promise of handling them quickly without giving up image understanding, language support or the usual safety protections, in his own words at the event.

The announcement, a secondary item on the Dev Day agenda, drew attention for its resemblance to Jev, the decision interface that startup TypeSafe AI launched earlier this month. TypeSafe is led by CEO Diogo Almeida, a former OpenAI engineer credited with co-inventing reinforcement learning techniques; the company did not respond to TechCrunch's questions about its former employer's new product, though Almeida joked on X about the start of the "clone wars."

The comparison between the two products is, for now, more conceptual than demonstrated. OpenAI has only opened limited access to the Decisions API and has not published latency, cost or accuracy figures that would allow Altman's speed claim to be tested. One data point does circulate about TypeSafe, though a partial one: at a recent hackathon, Shapor Naghibzadeh — a cybersecurity professional who ran the startup QueryStory — estimated that monitoring an agent's actions with Jev cost $2.94, compared with $372 using large state-of-the-art models. It is a one-off demonstration of a prototype, not an independent benchmark, and it should be read that way.

The most relevant use case is not classifying images but watching agents. OpenAI has acknowledged incidents in which its agents behaved improperly while operating on the open internet and has responded, among other measures, by using a separate model to detect problematic behavior — an approach that consumes a significant amount of compute. Naghibzadeh's approach with Jev points in the same direction, but more cheaply: check each agent action against the assigned task, block with high confidence whatever it flags as problematic, route doubtful cases to human review and let the rest through. It has been presented as a hypothesis that, in theory, could have prevented an incident at Hugging Face, though without a reconstruction to prove it; its real effectiveness depends on the model telling the legitimate from the dangerous, and on not over-blocking harmless tasks.

Almeida frames TypeSafe's bet in terms of the distinction, borrowed from cognitive psychology, between "System One" (fast, intuitive thinking) and "System Two" (deliberate reasoning): in his view, OpenAI's interest in a similar product suggests that building an artificial "System One" — cheap, fast, bounded — is a direction for the future. It is a reasonable sales argument, but not proof: not every software problem reduces to discrete choices, and a faster classifier does not replace broad reasoning when the task genuinely demands it.

Our read is that this announcement matters less for what it proves — so far, almost nothing that can be independently verified — than for what it signals. The contest between the big labs is no longer fought only on reasoning benchmark tables: it is also fought over who manages to get an agent to act, and above all to be monitored, for cents instead of hundreds of dollars. It is the same logic we have been seeing in the falling cost of cloud inference and in the rise of lightweight on-device models: frontier intelligence is reserved for the hard problems, and everything else is delegated to cheaper, specialized layers.

That move toward bounded, auditable decisions is, at bottom, one piece of the governance infrastructure that agentic AI needs in order to scale safely: the cheaper it becomes to review every action an agent takes, the easier it will be to monitor all of them instead of sampling a few at random. In the long run, that kind of control layer — combined with human oversight for doubtful cases, as Naghibzadeh's approach proposes — is part of what is needed for the promise of an economy of reliable agents, freeing human time for the work that really matters, not to remain keynote talk.

But in the short term, caution is the sensible stance. OpenAI has not published metrics of its own; the cost comparison in circulation belongs to a third party's prototype presented at a hackathon; and the near-simultaneous appearance of a product apparently similar to that of a smaller startup — led, moreover, by a former OpenAI engineer — raises the usual Silicon Valley question about how much room is left for a specialized competitor when the giant decides to enter its niche. If the Decisions API delivers what it promises, the winners are the developers who today overpay to use large models on simple tasks; if it fails to tell a correct decision from a cheap one, the savings will be paid for in errors that only surface once the product leaves preview.