OpenAI makes its Jev competitor widely available as decision model competition heats up

OpenAI opened its Decisions API to all customers on Thursday, following an announcement at its developer day on September 29. The product enters a market for fast, cheap classification-and-choice models that TypeSafe, a startup founded by a former researcher at the frontier lab, helped create with its Jev model. TypeSafe has reportedly reached a $7.6 billion valuation on the strength of that category.
The idea behind a decision model is that, given a set of options, it picks one and attaches a confidence rating. Developers adopted such models—sometimes called System One models—as substitutes for large language models in workflows where speed and cost matter more than open-ended generation. Rather than train a purpose-built decision model, OpenAI took ChatGPT Luna and reconfigured it to accelerate inference and lower cost, according to Nikunj Handa, the company's product lead for APIs. Handa said the team's reaction to Jev was that there were a million use cases for such a capability, and it moved quickly to ship a competitor.
The headline number for OpenAI is latency. Handa said the Decisions API cuts response times from one to three seconds down to roughly 100, 120 or 180 milliseconds. He described the current offering as reconfigured Luna rather than a specially calibrated decision model, leaving open the possibility of dedicated training later. In practice, that framing suggests OpenAI is leaning on serving optimizations—faster inference paths, geographic placement of compute—rather than a new architecture. The company also targeted inference resources in different regions to reduce round-trip latency for customers worldwide, and added voice and visual modalities to the public version that were not part of the initial announcement.
Early customers include Box, Datadog, Ramp and Speak. Andrew Hsu, CTO of Speak, which builds AI language tutors, said the API reduced the gap between actions from 1.4 seconds to 0.13 seconds. OpenAI also used the product internally to make its Dots voice assistant more responsive, a natural fit for conversational software where delays stretch perceived response time.
Performance claims across the market are contested. TypeSafe CEO Diogo Almeida has argued against trusting benchmarks and urged teams to test models on their own workloads, even as TypeSafe published its own rankings last week alongside a new Jev release adding support for images and smart-if statements. Public indices—the AIM System One Index and the Artificial Analysis Decision Index—currently favor Liquid's D1 Ultra, released last week, and Base Lab's Elephant-1. The latest Jev places seventh and third on those two indices, while OpenAI's decision model places eighth and fourteenth. On OpenRouter, TypeSafe's models remain the most widely queried. There is also JevMan, a benchmark in which models play PacMan, where the latest Jev leads; Felix Wunderlich, chief growth officer at benchmark maker Opper AI, said the game is a useful test because it demands a continuous stream of quick, high-stakes choices every few hundred milliseconds.
OpenAI's likely edge, based on the reported details, is multimodality. Handa said Luna is strong at image classification and has shown better calibration there, and that customers want Decisions API to be more flexible about returning images as outputs as well. A model that can both consume and emit visual content can serve workflows—screening, tagging, routing camera frames—that a text-only classifier cannot touch. That is a plausible differentiator against Jev, though the index standings suggest it has not translated into leadership on general decision benchmarks.
Speed remains the frontier. Handa raised the prospect of pushing latency toward 60 milliseconds, which would open calls on essentially every frame of a video. This is my own reading, but that threshold matters because it crosses from interactive use into real-time perception, where a decision model could sit inside a live pipeline rather than answer a discrete request. The unit economics follow the same logic: a classifier that is cheap and fast enough can be applied far more liberally than an LLM, the kind of efficiency gain that tends to make a resource more ubiquitous rather than merely cheaper.
A practical trade-off is calibration. A confidence rating is only useful if it correlates with correctness, and calibration can drift when a general-purpose model is repurposed for narrow choices. OpenAI's own framing—plain Luna tuned for speed—implies calibration quality may vary by task, which is consistent with Almeida's advice to validate on your own data. Teams replacing an LLM call with Decisions API should also watch edge cases: ambiguous options, ties and out-of-distribution inputs, where a fast classifier can fail more quietly than a slower generative model that can explain itself.
Why it matters: Developers building latency-sensitive features—voice agents, moderation, routing, visual tagging—now have a first-party OpenAI option that trades model generality for speed and price, and a crowded set of rivals in Jev, D1 Ultra and Elephant-1 to compare against. The practical work shifts from prompt engineering toward benchmarking and calibration on your own workload, since public leaderboards disagree and vendor rankings are self-published. For product teams, sub-200-millisecond decisions make per-interaction or per-frame calls economically thinkable; whether OpenAI closes the gap on the indices, or opens a lead through multimodality, is likely to be decided by customer trials rather than benchmark tables.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
OpenAI is launching its decision model to the world. Can it compete with the original Jev?