OpenAI Opens the Cheap-Decision Race

The practical claim is speed. OpenAI’s API product lead, Nikunj Handa, told TechCrunch the company reworked its Luna model for faster, cheaper inference, shrinking response times from roughly one to three seconds to about 100 to 180 milliseconds. That matters because plenty of software jobs do not require a lyrical answer. They require a fast classification, a ranked option, or a confidence score. In product terms, that is the difference between something that feels embedded and something that feels bolted on.

OpenAI had announced the Decisions API at its developer day on September 29. TechCrunch reports that the version now available to all customers adds voice and visual modalities, while OpenAI also worked on reducing latency across regions by targeting inference resources geographically.

The category itself owes something to TypeSafe’s Jev model, which helped turn “decision models” into a recognizable business rather than just a technical trick. As TechCrunch notes, Jev showed developers that a super-classifier could stand in for a large language model across some workflows, operating faster and at much lower cost. That is not a philosophical shift so much as a budgetary one. If a task mostly needs yes, no, route, sort, or label, verbosity becomes overhead.

Handa told TechCrunch that OpenAI’s team saw Jev and immediately recognized “a million use cases” for the approach. Instead of building a specially trained decision model from scratch, OpenAI used Luna and tuned it for speed. There is a slightly unromantic quality to that admission, but it may also be the point. AI competition has a habit of sounding abstract until a company says, in effect, we made it quicker and cheaper and that was enough.

Early customer examples suggest the appeal is throughput, not mystique. TechCrunch reports that Speak, which builds AI language tutors, cut the time between actions from 1.4 seconds to 0.13 seconds using the Decisions API. OpenAI also used the product to improve the responsiveness of its Dots voice assistant.

What this does not establish is category dominance. TechCrunch cites public benchmarks where Liquid’s D1 Ultra and Base Lab’s Elephant-1 tend to rank ahead, with OpenAI’s decision model trailing Jev on some indices. That is awkward, but hardly disqualifying in a market where buyers may care less about leaderboard status than about latency, cost, image handling, and whether a model behaves predictably inside an existing stack.

OpenAI’s own pitch seems to acknowledge that. Handa pointed to Luna’s multimodal strengths, especially for image classification, and floated still faster response times as a path to new uses, including video-related workloads. For now, though, the news is simpler than the ambition: OpenAI saw a useful specialist market taking shape and decided not to leave it to specialists.