The OpenAI Decisions API sprint
Nikunj Handa runs OpenAI’s APIs, including the new OpenAI Decisions API. When his colleagues first looked at TypeSafe’s Jev — the decision model that started acting like a super classifier, taking a set of options and returning a choice plus a confidence rating — they saw demand worth chasing.
“Wow, we have like a million use cases for this,” Handa told TechCrunch. His team sprinted to offer its own version. This week, that version went widely available: the Decisions API, open to all OpenAI customers, announced at the company’s DevDay on September 29, through public beta on October 6, to general availability on October 22 — a little over three weeks from first preview to launch.
Here is what makes that sprint worth watching. OpenAI did not build a new kind of model. On Handa’s account, rather than building a specifically calibrated decision model, the company took ChatGPT Luna and reconfigured it to speed up inference and lower the cost. Handa was cheerfully blunt about it: “The big unlock with Decisions API has been the speed, shrinking things from one to three seconds to 100, 120, 180 milliseconds. In the future, we may sort of like have specific training to make our models like even better in these ways. But right now, it’s just plain old Luna that gets the job done.”
Read that again: one of the industry’s most capable labs entered a brand-new model category by reconfiguring an existing flagship rather than building a purpose-built decision model.
Why the hurry
The hurry makes sense once you see the market. TypeSafe, founded by a former researcher at a frontier lab, recently became a $7.6 billion company (per TechCrunch; the October 9 raise that did it was reported as $870 million at a $7.5 billion valuation, led by Andreessen Horowitz). Jev created the category: software developers discovered that these decision models — System One models, named for fast, intuitive thinking — could replace large language models in a variety of workflows, operating more quickly and at dramatically lower cost.
So OpenAI’s urgency may be less the story of a frontier lab inventing than of a platform owner reacting to a demand explosion in a category it did not create. The Decisions API is priced at $0.10 per million uncached input tokens, with no output-token charge at all — the published pricing also includes discounted cache hits and charged cache writes, conditions that could shift the real-world picture. A developer comparison puts Jev cheaper on raw input ($0.042 per million), with an illustrative million 500-token classifications costing about $21 on Jev versus $50 on the Decisions API at the cited base input rates. OpenAI is not undercutting Jev on price — whether that is a land grab or simply how a platform prices is an open question.
The mid-table problem
And here is the delicious part: by the benchmarks’ own telling, OpenAI’s model trails the named leaders on both indices. Public indices — the AIM System One Index and the Artificial Analysis Decision Index — tend to favor Liquid’s D1 Ultra and Base Lab’s Elephant-1; the latest Jev ranks seventh and third; OpenAI’s Decision Model falls eighth and fourteenth.
Eighth and fourteenth. A company of OpenAI’s resources is ranking mid-table — and it launched anyway. That is either reckless or instructive. It is instructive, because the benchmarks may not be measuring the thing the market is buying.
TypeSafe CEO Diogo Almeida distrusts benchmarks and urges people to trial the models in their own work. Yet TypeSafe’s offerings are, for now, the most widely-queried decision models on OpenRouter. The market is voting with traffic, not leaderboards. And nobody can quite say what makes a good decision model — “it works for you” is, honestly, the best definition anyone has. Optimization across intelligence, cost, calibration, and speed differs by workload.
OpenAI is, in other words, shipping into a market where the scoreboard is unsettled and the demand is obvious. The steelman case for its strategy is strong: OpenAI already reaches the developers and operates inference at global scale, and Luna brings multimodality — voice and visual inputs joined at general availability — a potential advantage, since multimodal inputs open use cases text-only models cannot touch. “Luna is extremely powerful and has been showing better calibrations,” Handa said of image classification. The reconfigured flagship may simply be good enough, at reported speeds of 100–180 milliseconds, for the customers already on the platform.
The customers are already trying
The early evidence for “good enough” is usage-based, not benchmark-shaped. In initial trials, Box, Datadog, Ramp, and Speak used the API to replace OpenAI’s own LLMs with higher throughput. Andrew Hsu, the CTO at Speak, said the Decisions API allowed the company to cut the time between actions from 1.4 seconds to 0.13 seconds. OpenAI itself used the product to improve its Dots voice assistant’s responsiveness.
Note the detail that usually gets buried: these customers replaced OpenAI’s LLMs. If selling a cheaper, faster product that substitutes for your own flagship is self-disruption, OpenAI is doing it cheerfully and in public — or is it simply following the demand? This is the part of the story no benchmark captures, and the question it leaves open is the interesting one.
OpenAI added new voice and visual modalities to the public version, and worked to ensure lower latency worldwide by targeting its inference resources geographically — the unglamorous infrastructure work that is, in a latency-sensitive category, the entire game. Handa is already thinking further ahead: could this reach 60 milliseconds — something like an API call for every single frame of a video?
The Jevons trap
There is a wry footnote here, and it sits in the article’s own kicker. TechCrunch’s kicker invokes the Jevons paradox — the idea that using a resource more efficiently makes it more ubiquitous. Could the same hold for automated decisions: make them radically cheaper and faster, and they start happening everywhere, in places nobody thought to automate? Handa sees an expanding set of use cases hiding in plain sight — framed as a question, that is a tantalizing prospect, not a proven one.
That is the real affair in this story. The benchmark crown, the purpose-built training, the calibration scores — these are the things the technical world argues about. But the market that just handed a startup an $870 million round on the thesis that agentic pipelines need a decision layer is not waiting for the argument to settle. OpenAI looked at TypeSafe, saw a million use cases, and shipped plain old Luna.
The question is not whether OpenAI built the best decision model. It is whether, in a category where nobody agrees what “best” means, combining real distribution with low-latency inference and simple pricing is the same thing as winning. The next time you see a decision model answer in about 100 milliseconds, remember: it did not arrive because someone solved intelligence. It arrived because someone saw the demand and sprinted.
\n
