On October 22, OpenAI opened Decisions API to all customers in general availability (GA). With this move, the company entered the "decision models" market directly — a direction pioneered by TypeSafe's Jev model, where large language models resolve small but frequently repeated choices in workflows in milliseconds rather than seconds.
Decisions API is built on GPT-6 Luna and, according to the company, runs up to 10 times faster than GPT-6 Luna via the Responses API. The public version adds voice and vision modes, and inference runs on geographically targeted resources.
How Decisions API Works
The API lets developers define a question and a set of allowed answers; context is supplied as text, images, or local JSON. The service returns three response types: the probability that a given criterion is met (probability), a selection from predefined options (choice), and a rubric-based score (score).
Up to 400 questions can be asked in a single request — twice the previous limit. The API supports including images in question prompts and prompt caching for repeated context; inference residency is available in the United States and the European Union. Details are in OpenAI's official announcement.
Pricing is strictly simplified: $0.10 per 1 million input tokens, $0.01 for cached input; output tokens are not charged at all. This pricing dramatically lowers the cost of high-frequency calls — exactly what made decision models an alternative to traditional LLMs.
Speed and Pricing: How It Was Achieved
According to TechCrunch, instead of building a separately trained decision model, OpenAI retuned ChatGPT Luna for speed and affordability. API product lead Nikunj Handa says latency dropped from one to three seconds to 100–180 milliseconds.
"The key breakthrough in Decisions API is speed: cutting from one to three seconds down to 100, 120, 180 milliseconds," Handa said. "In the future, we may specifically train our models to make them even better in this direction. But for now, the work is simply done by an ordinary Luna."
He added that his team, upon seeing Jev, concluded "we have millions of use cases for this" and quickly built their own version. The public version adds voice and vision modes; additional capacity has been brought online in several regions of Asia.
Decisions API is also now natively integrated with GPT-Live: voice apps can make instant decisions mid-conversation — for example, a customer support agent chooses between asking a clarifying question, routing the request, or escalating the issue.
The service was announced on September 29 at the company's developer day and has been tested in beta since. After going public, Handa also revealed a goal of even greater speed: "Can we bring this down to roughly 60 milliseconds — then you could imagine one API call per video frame," he told TechCrunch. In image classification, he said, "Luna is very strong and shows better calibration," while customers want to use Decisions API more flexibly — including with images as output.
Early Users and the Competitive Landscape
In early tests, engineers at customers such as Box, Datadog, Ramp, and Speak replaced OpenAI's LLMs with high-throughput decision calls. According to Speak CTO Andrew Hsu, whose company builds a language-learning app, the API cut the time between actions from 1.4 seconds to 0.13 seconds. OpenAI itself also used the product to speed up its Dots voice assistant.
Internal adoption is broad too: the Personal Finance team uses Decisions API to categorize transaction descriptions, while the Communications team uses it to classify media coverage by topic and whether it reinforces or weakens key narratives.
The competitive environment is heating up. Public benchmarks — the AIM System One Index and the Artificial Analysis Decision Index — currently rate Liquid's D1 Ultra and Base Lab's Elephant-1 highly; the latest Jev model ranks seventh and third in these two indexes respectively, while OpenAI's decision model ranks eighth and fourteenth. At the same time, according to OpenRouter data, the Jev family remains the most requested decision models.
TypeSafe released the latest version of Jev last week — with image support and smart-if expressions. Previous coverage of the company is available in the site archive.
One of the most interesting tests of decision models is the JevMan benchmark highlighted by TechCrunch: in this test, compiled by Opper AI, models play the game PacMan. According to Opper AI growth director Felix Wunderlich, the game is an interesting test because it is a "continuous, fast, and important stream of decisions" — at every turn the model has only a few hundred milliseconds to choose where to go. In this ranking, Jev's latest model currently holds first place.
