A customer support system might decide whether to refund an order, send a case to a human or request more information. An agent might choose which tool to call next. A security platform could classify a login attempt as safe, suspicious or blocked. In each case, the useful output is not a fluent response. It is a structured choice that another system can act on immediately.
OpenAI is now packaging that behavior as a product. In its announcement, OpenAI said it had made the Decisions API generally available, describing it as an endpoint based on its GPT-6 Luna model for bounded, typed decisions from text, images and structured inputs.
Read The Next AI Bottleneck Is Decision Making, Not Generation
The distinction matters because conventional language models often spend time generating output that software must then interpret. A decision model returns an answer in a predefined format, such as a label, Boolean value, score or selection from a list. That can make the model easier to connect to automated systems and less costly to run at high volume.
OpenAI says the Decisions API can deliver latency up to 10 times lower than its Responses API. Its announced input price is $0.10 per million tokens. The company’s positioning is aimed at workflows where a delay of a second or two can make an agent feel sluggish, especially when several decisions must happen in sequence.
The practical effect could be visible in small moments. A shopping assistant may decide which inventory service to query before a user notices the pause. A workplace agent could determine whether a request needs approval while the employee is still typing. Developers may begin treating AI decisions like fast infrastructure, placing them between events, tools and people rather than presenting them as a destination for conversation.
TypeSafe is pursuing the same opportunity with a different framing. In its introduction to TypeSafe’s founder Diogo Almeida described Jev as the company’s first System One model. Jev is designed to return typed probabilistic decisions rather than generated text. TypeSafe claims that it can deliver performance comparable to large language models on decision tasks while operating with much greater speed and efficiency.
That probabilistic output is central to the product category. A system can choose “approve,” “reject” or “review,” but it can also indicate how strongly it favors each option. In a workflow, those scores might determine whether an automated action proceeds or whether a human should inspect the case. The confidence estimate becomes part of the control surface, not merely an internal model statistic.
Vercel has helped make Jev visible to developers. Vercel reported that Jev reached more paid teams in its first 24 hours than any previous model in its AI Gateway. The company described use cases involving parallel typed choices, scores and Boolean decisions in software workflows.
That adoption claim does not establish that Jev is the best system for every task. It does suggest that developers recognize a gap between the models used for open ended generation and the smaller, sharper tools needed to coordinate applications. When a workflow contains dozens of tiny judgments, a model built for long answers may be excessive.
The competitive test will therefore involve more than intelligence in the familiar sense. Buyers will need to examine calibration, or whether a model’s stated confidence matches its actual reliability. A system that is fast but confidently wrong can create more operational risk than a slower model that knows when to defer.
Multimodal input will also shape the choice. OpenAI says its Decisions API can process text, images and structured inputs, which could allow one decision layer to inspect documents, screenshots and database fields. TypeSafe’s announcement emphasizes typed probabilistic outputs, but the sources provided here do not establish an equivalent multimodal scope for Jev. That difference may matter in businesses where decisions begin with visual or mixed data.
Cost is another variable. A low per token price is useful when a company makes millions of small calls, but the full expense also includes retries, monitoring, human review and the engineering work required to define safe option sets. A decision system can be cheaper than a general model only when the surrounding workflow is designed to use its narrow strengths.
There is some independent evidence about Jev, although no comparable opposing assessment was found in the supplied sources. Researchers writing in an independent evaluation of Jev examined it across 37 datasets and reported that its choice probabilities were well calibrated, while comparing its performance with open weight language models. The paper is an outside assessment, not a company announcement, but it does not settle how Jev compares with OpenAI’s system across real production workloads.
That leaves the larger question open. Are decision models a durable category, or are they mainly a cheaper configuration of existing language models? The answer may depend on whether developers start designing products around them.
If bounded choices become a standard layer in applications, users may rarely see the model itself. They will experience faster approvals, more responsive agents and fewer pauses between one automated action and the next. The intelligence will be less like a chatbot waiting for a prompt and more like an invisible network of traffic signals, constantly directing software toward its next move.
For OpenAI and TypeSafe, that future turns speed into a product philosophy. The winner will not necessarily be the model that writes the most convincing answer. It may be the one that makes the right small decision, reveals how certain it is and gets out of the way before anyone notices it was there.
Maya Lindqvist is not a person. No notebook, no deadlines, no face behind the name — just a byline this newsroom publishes under. Here is the production line underneath it, because a name beside a portrait reads like a journalist, and this one is not one.
The models. Writing: gpt-5.6-luna and qwen3-max. Out on the live web: gpt-5.6-luna and gpt-5.6-terra. Pictures: gpt-image-1 and gpt-image-1-mini. Swap one in the newsroom and this line swaps with it — it is read off the machines, not typed here.
How a story is made
- Research. The searching model reads around the story, pointed at primary sources — the filing, the post, the repository — rather than at somebody else's write-up of them.
- Writing. The writing model drafts it against what was found, at Maya Lindqvist's usual length and in Maya Lindqvist's usual register.
- The loop. A reviewer reads the draft and sends it back with notes. Then reads it again. A piece can go round several times before it leaves the building.
- Enrichment. A quotation has to appear word for word on the page it is taken from. A chart may only use figures that appear in the source it cites. Whatever fails is dropped, and the reason is kept.
- Fact check. A last pass hunts for claims the article makes and its sources do not.
- A human stop. Sensitive subjects are held for a person to read before publication, and a person can kill any of it at any point.
If that sounds less like a newsroom and more like a factory: quite. It is called Press Factory.
This article was generated using AI and published automatically without human pre-publication review.
How this article was made
The article was produced by the Grandmonts Media News Engine using automated research, drafting and verification workflows. No human editor reviewed the article before publication. Grandmonts Media remains responsible for the published content. Errors can be reported at office@grandmonts.cz.