OpenAI makes Decisions API widely available

OpenAI has opened its Decisions API to all customers, bringing its version of a fast, low-cost decision model to developers. The company announced the API at its developer day on September 29; it became widely available on Thursday.

The product follows the rise of decision models such as TypeSafe’s Jev, which can choose among a set of options and provide a confidence rating. Developers have found that these models can replace large language models in some workflows, making decisions faster and at dramatically lower cost.

OpenAI’s approach was not to build a specially calibrated decision model. Instead, the team adapted ChatGPT Luna, reconfiguring it to reduce inference time and cost. Nikunj Handa, the product lead for OpenAI’s APIs, said the team saw many potential uses for this kind of model after Jev emerged.

Speed was the main improvement, Handa said: Decisions API responses that previously took one to three seconds can now arrive in 100 to 180 milliseconds. He said OpenAI may add specific training in the future, but for now, “plain old Luna” is doing the job. The public version also includes voice and visual modalities, while OpenAI has targeted inference resources geographically to reduce latency in regions around the world.

Faster decisions show value in customer trials

OpenAI says the Decisions API’s main advantage is speed: by reconfiguring ChatGPT Luna to accelerate inference and lower costs, it reduced response times from one to three seconds to roughly 100 to 180 milliseconds. In initial customer trials, engineers at Box, Datadog, Ramp, and Speak used the API to replace OpenAI’s LLMs with a higher-throughput option.

For Speak, which builds AI language tutors, the change cut the time between actions from 1.4 seconds to 0.13 seconds, according to CTO Andrew Hsu. OpenAI also used the API to make its Dots voice assistant more responsive, offering an early example of how faster decisions can improve voice interactions.

The public release adds voice and visual modalities, extending the API beyond text-based decisions. OpenAI also targeted inference resources geographically to help keep latency low in regions around the world. Together, these changes aim to make fast decision-making available across more customer workflows and locations, while the trials provide practical evidence of the API’s potential benefits.

Decision-model benchmarks show a crowded field

There is no universal winner among decision models: the right balance of intelligence, cost, calibration, and speed depends on the workload. TypeSafe CEO Diogo Almeida cautions against relying on benchmarks and recommends trying models on the work they need to do. Still, TypeSafe published rankings alongside its latest Jev release, which adds image support and smart-if statements.

On public benchmarks, Liquid’s D1 Ultra and Base Lab’s Elephant-1 tend to lead the AIM System One Index and Artificial Analysis Decision Index. Jev remains competitive, placing seventh and third, respectively, while OpenAI’s Decision Model ranks eighth and fourteenth. TypeSafe’s models are currently the most widely queried on OpenRouter.

Jev’s results also vary by test: its latest model tops JevMan, where decision models play PacMan. The benchmark’s creator says the game tests a continuous stream of quick, high-stakes choices, with a decision needed at each junction within a few hundred milliseconds. These differing standings underline why developers may need to judge models against their own priorities rather than assume one ranking settles the question.

Multimodality and speed could unlock more uses

OpenAI may have an edge in multimodality. For image classification, Luna is “extremely powerful” and has shown better calibration, Handa said. Customers are also asking for the Decisions API to be more flexible about using images as outputs, pointing to interest in models that can handle visual inputs and produce visual results.

Handa sees further gains in speed opening up applications that are difficult to support with slower responses. He asked whether the API could reach about 60 milliseconds, making it possible to call the model for every frame in a video. That would extend decision models beyond selecting among options in a single interaction to processing a continuous stream of visual information.

The current API has already reduced response times from seconds to roughly 100–180 milliseconds, according to Handa. As decision models become faster and cheaper, more workflows may become practical. Handa expects an expanding “crazy set of use cases” to emerge—applications that had been hiding in plain sight until decision models made them efficient enough to use.