OpenAI opens Decisions API as decision-model race accelerates

A faster version of Luna

The category took shape around Jev, TypeSafe’s model for selecting an option and estimating confidence. Developers saw that these decision models, also called System 1 models, could replace large language models in some workflows while running faster and costing less.

OpenAI moved quickly after its team saw Jev, Nikhunj Khanda, the company’s API product lead, told TechCrunch. Rather than build a specially calibrated decision model, the team adapted ChatGPT Luna to speed up inference and lower costs. The API was announced at OpenAI’s developer day on September 29 and opened to all customers on Thursday.

The public release adds voice and visual inputs. OpenAI has also placed inference resources in different regions to reduce latency. For now, the API still uses the general-purpose Luna model; Khanda said OpenAI may train models specifically for decision tasks later.

Early results, uncertain rankings

Engineers at Box, Datadog, Ramp and Speak tested the API as a replacement for OpenAI’s large language models, aiming to increase throughput. At Speak, which builds AI language tutors, CTO Andrew Xu said the time between actions fell from 1.4 seconds to 0.13 seconds. OpenAI also used the product to improve the response speed of its Dots voice assistant.

The gains are clear in these examples; the broader question of how to judge decision models is not. Performance depends on the task and the trade-off between reasoning quality, cost, calibration and speed. TypeSafe CEO Diogo Almeida is skeptical of benchmarks and recommends testing models on a company’s own work. Yet TypeSafe published rankings last week alongside a new Jev version, adding image support and intelligent conditional operators.

On the public AIM System One Index and Artificial Analysis Decision Index, Liquid’s D1 Ultra and Base Lab’s Elephant-1 lead. The new Jev ranks seventh and third, while OpenAI’s Decision Model ranks eighth and fourteenth. TypeSafe models are still most often accessed through OpenRouter.

JevMan offers a less conventional test: models play Pac-Man. The latest Jev ranks first. Felix Wunderlich, growth director at benchmark creator Opper AI, said the game tests a stream of fast, high-stakes decisions, with a direction to choose at each turn in a few hundred milliseconds.

The question beyond speed

I think the most revealing detail is that OpenAI’s current product is an adaptation of Luna, not a purpose-built decision model. That makes the latency improvement meaningful, but leaves open how much comes from the model itself and how much from the way the API is tuned for a narrower job.

Khanda said Luna is particularly effective at image classification and has better calibration, and that OpenAI customers want more flexibility for image outputs. He also sees further speed gains as a way to expand where the API can be used: at around 60 milliseconds, it might be possible to call it for every video frame.

The announcement is quiet about how performance holds up across customers’ own tasks, especially when speed competes with reasoning quality and calibration. That matters because lower cost and latency could make decision models useful in far more places. If that happens, the test will not be whether they beat a benchmark, but whether the savings survive once developers put them into real workflows.