OpenAI has opened broad access to its Decisions API, a decision-making model family announced at the company’s developer day on September 29, 2024. The API became available to all customers on the Thursday following that announcement. The move positions OpenAI in the emerging market for fast, low-cost “System One” decision models that select among options and report confidence scores.
How the model was built and what changed
Rather than training a brand-new, purpose-built decision model, OpenAI’s API team adapted ChatGPT Luna to reduce inference time and cost. Nikunj Handa, product lead for OpenAI’s APIs, told TechCrunch that the team responded quickly after seeing TypeSafe’s Jev model and its wide applicability.
Handa said the main benefit has been speed: response times that were one to three seconds dropped to roughly 100–120–180 milliseconds. He also noted the possibility of future specialized training to further optimize decision performance, but that the reconfigured Luna already meets many needs. The public release adds new voice and visual modalities, and OpenAI has tried to lower latency worldwide by routing inference resources geographically.
Early customer results
In pilot deployments with customers such as Box, Datadog, Ramp, and Speak, engineers used the Decisions API to replace higher-latency LLM calls with higher-throughput decision models. Andrew Hsu, CTO of Speak (which builds AI language tutors), reported that the Decisions API reduced time between actions from 1.4 seconds to 0.13 seconds. OpenAI also used the product to make its Dots voice assistant more responsive.
Handa suggested that further speed gains could enable new use cases: for example, reaching around 60 milliseconds could make it feasible to call the API for every frame in a video stream.
Competitive landscape and benchmarks
TypeSafe — a startup whose Jev model created the category of fast decision models — has driven significant market attention; it was previously described as a $7.6 billion company. TypeSafe recently released a new Jev model that adds image support and "smart-if" statements, and it published its own rankings. Diogo Almeida, TypeSafe’s CEO, has expressed skepticism about benchmarks and recommends users test models in their own workloads.
Public benchmarks such as the AIM System One Index and the Artificial Analysis Decision Index show varied results. These indexes have tended to favor Liquid’s new D1 Ultra and Base Lab’s Elephant-1. On the latest published indices, TypeSafe’s new Jev model ranks seventh and third respectively, while OpenAI’s Decision Model ranks eighth and fourteenth. Separately, TypeSafe’s models are currently among the most queried on OpenRouter.
A more playful benchmark, JevMan, tests decision models by having them play Pac-Man; the most recent Jev model tops that ranking. Felix Wunderlich, chief growth officer at Opper AI (the benchmark’s creator), said Pac-Man is useful because it forces a nonstop stream of fast, high-stakes decisions, where the model must choose directions in a few hundred milliseconds.
Multimodality, calibration and what to watch
OpenAI’s potential advantage may lie in multimodal capabilities. Handa told TechCrunch that Luna is “extremely powerful” for image classification and has shown strong calibration, and that customers want greater flexibility to use images as outputs of the Decisions API.
There is no single objective measure that defines a "good" decision model; workload-specific trade-offs among intelligence, cost, calibration and speed mean the right choice varies by use case. As these models become faster and cheaper, they are likely to enable new and previously overlooked applications — a trend OpenAI’s team expects will reveal a wide set of fresh use cases.
