OpenAI opens its Decisions API to all customers as decision-model race with TypeSafe heats up

OpenAI has made its Decisions API available to all customers for the first time, according to TechCrunch, formally entering the fast-growing market for decision models: lightweight systems that make rapid choices at a small fraction of the cost of a large language model.

A category a rival built

The market OpenAI is moving into was largely defined by TypeSafe, a startup founded by a former researcher at a frontier AI lab that TechCrunch reports is now a $7.6 billion company. Its Jev model popularised a new class of deep learning systems sometimes called super classifiers or System One models: presented with a set of options, they pick one and attach a confidence score. Developers found these models could stand in for LLMs across many workflows while responding faster and at dramatically lower cost.

Nikunj Handa, product lead for OpenAI's APIs, told TechCrunch that seeing Jev prompted his colleagues to conclude there were countless potential applications for the approach. Rather than train a purpose-built decision model, OpenAI took ChatGPT Luna and reconfigured it to accelerate inference and cut cost. Handa said the big gain has been latency, with responses shrinking from the one-to-three-second range typical of LLM calls to roughly 100 to 180 milliseconds, and that more specialised training may come later.

What shipped

The API was announced at OpenAI's developer day on September 29 and became available to every customer on Thursday, TechCrunch reports. The public release adds voice and visual modalities, and OpenAI positioned its inference resources geographically to keep latency low in regions around the world.

In initial trials, engineers at Box, Datadog, Ramp and Speak used the API to replace OpenAI's LLMs where throughput mattered. Andrew Hsu, CTO of language-tutoring company Speak, told TechCrunch the switch cut the time between actions from 1.4 seconds to 0.13 seconds. OpenAI also used the product internally to make its Dots voice assistant more responsive.

Crowded field, murky benchmarks

What makes a good decision model remains unsettled, since the trade-offs between intelligence, cost, calibration and speed appear to differ by workload. TypeSafe CEO Diogo Almeida distrusts benchmarks and urges buyers to test models on their own tasks, though TechCrunch notes TypeSafe published its own rankings alongside its latest Jev release, which added image support and smart-if statements.

Public leaderboards such as the AIM System One Index and the Artificial Analysis Decision Index currently favour D1 Ultra, a new decision model from Liquid, and Base Lab's Elephant-1. The newest Jev sits seventh and third on those two indices, while OpenAI's entry lands eighth and fourteenth. TypeSafe's models remain the most queried on OpenRouter, according to TechCrunch.

There is also JevMan, a benchmark from Opper AI in which decision models play Pac-Man, a game demanding a nonstop stream of quick, high-stakes choices; the latest Jev tops that ranking. Opper chief growth officer Felix Wunderlich told TechCrunch the game is a useful test because every junction forces a decision within a few hundred milliseconds.

Where OpenAI sees an edge

Handa pointed to multimodality as a likely advantage, telling TechCrunch that Luna performs strongly on image classification and shows better calibration there, and that customers also want the API to be more flexible about producing images as outputs. He also sketched a faster future: at around 60 milliseconds per call, he suggested, an API request could run on every single frame of a video.

Why it matters

An OpenAI entrance is the strongest signal yet that decision models have moved from startup curiosity to mainstream infrastructure. A category invented by one company has grown valuable enough to pull in the industry's largest vendor, and OpenAI's approach of reusing an existing model shows how quickly incumbents can commoditise a new niche. For developers, the economics are the real story: swapping second-long LLM calls for millisecond-scale decisions changes what is practical in real-time products, from voice assistants to agents to per-frame video processing. There is also a Jevons-style dynamic at work, the source notes, in which making a resource cheaper to use tends to make it far more ubiquitous, and Handa expects an expanding set of applications that were hiding in plain sight. With benchmarks disagreeing on leaders, buyers will need to evaluate these models on their own workloads rather than trust any single ranking.