At DevDay, OpenAI unveiled the Decisions API, built on the GPT-6 Luna model, which can quickly choose from preset answers and return confidence scores with only 150ms latency. This API closely overlaps with the core capabilities of startup Jev, which launched on September 15 and is valued at over $10 billion. The move could pose a significant competitive threat to Jev, sparking community discussion.
SynthePulse Insight · AI deep readingMembers
Version 1 · 1 source
At DevDay, OpenAI unveiled the Decisions API built on GPT-6 Luna, narrowing open-ended generation into "choosing from preset options and returning a confidence score," with a nominal latency of 150 milliseconds. This lands almost squarely on the core capability of Jev, which launched on September 15. But the limited preview, undisclosed pricing and tuning capabilities, and the hardest variable to test—just how accurate the probabilities really are—make "Jev is dead" look more like a judgment awaiting verification than a conclusion.
On September 15, TypeSafe AI officially launched Jev. Half a month later, at DevDay, OpenAI announced the new Decisions API: built on GPT-6 Luna, it can choose among a set of preset answers and return a corresponding confidence score. In terms of product form, both point to the same thing—developers provide the question, candidate answers, and context, and the model does not generate a passage of text but instead gives a decision that can be handed directly to a program.
This overlap is not a vague similarity. According to TypeSafe's official definition, Jev is a "System One Model" that returns structured results through three primitives—Choice, Score, and Noul—of which Choice selects one from up to 255 preset options and returns a probability distribution. And when introducing the Decisions API in the keynote, Sam Altman said it lets the model respond in a fraction of a second, by providing a set of preset options to the Luna model.
In other words, OpenAI has narrowed open-ended generation into "choosing from explicit options," which is highly consistent in design approach with Jev's Choice primitive. What is truly worth asking is not "how similar is it," but: when the platform player enters the field itself, how much time and space does a product that has been live for only 15 days have left?
OpenAI cites a latency of 150 milliseconds for the Decisions API, whereas using GPT-6 Luna directly for the same task takes 1.6 seconds. This comparison is the most concrete and most developer-perceptible number in this launch: it shows that the value of the Decisions API is not just "a different interface," but that narrowing the same underlying model's capability to a selection task buys roughly an order-of-magnitude difference in response.
Choosing Luna as the underlying model is also informative. It is the smallest and cheapest model in OpenAI's current lineup; using the smallest, lowest-cost model to handle decision tasks, rather than calling Astra or GPT-6.1 Sol, shows that this capability is clearly aimed at high-frequency, low-cost foundational components that can be embedded into agent workflows at scale.
What needs to be distinguished is this: the two numbers, 150 milliseconds and 1.6 seconds, come from OpenAI's claims and are source claims, not third-party independently verified results. They can illustrate product positioning, but they cannot be directly equated with end-to-end performance in a real production environment.
This article is based on a single secondary reporting source (AI Qianxian, source_tier 2) and the Reddit posts, Arcturus Labs blog, and Latent Space interview content it relays. The performance numbers from both OpenAI and TypeSafe (150 milliseconds, 1.6 seconds, 20 to 200 times faster, 1/40 to 1/400 the cost, $0.042 per million input tokens) are all claims from their respective sources and have not been independently verified; the Reddit user's internal test lacks verifiable methods and data and is presented only as a source claim. Jev's complete architecture and specific implementation have not been disclosed, so its technical path cannot be directly verified; the Decisions API's pricing, cap on candidate answers, and custom tuning capability have all not been announced. Judgments in the article such as "rushed launch," "architecture does not constitute a moat," and "window of opportunity narrows" are all inferences based on sources, not confirmed facts.
Primary report
Primary source