OpenAI makes its Jev competitor widely available as decision model competition heats up | TechCrunch
Short summary
OpenAI has made its Decisions API generally available, a fast, cheap model for picking among options with a confidence score, meant to compete with TypeSafe's Jev. The pitch is that for many workflows, a quick classifier-style model can replace a full LLM.
- OpenAI didn't build a dedicated decision model; it repurposed its ChatGPT Luna model and tuned it for lower latency and cost.
- Speed is the main selling point: response times dropped from 1-3 seconds to roughly 100-180 milliseconds, and Speak cut its action gap from 1.4 seconds to 0.13 seconds.
- Benchmarks don't settle which decision model is best. Jev ranks well on some indices and tops the PacMan-based JevMan test, while OpenAI's model ranks lower, though it may have an edge in image classification.
Full summary
OpenAI has made its Decisions API generally available, entering a fast-growing category of "decision models" that pick from a set of options and attach a confidence score. The category was effectively created by TypeSafe, whose Jev model led developers to see that these fast, cheap classifiers could stand in for large language models in many workflows. TypeSafe is now valued at $7.6 billion. OpenAI's version is not a purpose-built decision model. Its product lead, Nikunj Handa, says the team took the existing ChatGPT Luna model and tuned it for speed and cost, and the main payoff so far is latency. Response times dropped from one to three seconds to roughly 100 to 180 milliseconds.
The early results from customers are the most concrete evidence in the piece. Speak's CTO, Andrew Hsu, said the API cut the gap between actions from 1.4 seconds to 0.13 seconds, which matters for a language tutor that has to respond in conversation. Box, Datadog, Ramp, and OpenAI's own Dots voice assistant were also in initial trials. The public version added voice and visual inputs and moved inference capacity closer to users in different regions to reduce lag. Handa floated the possibility of pushing latency down to around 60 milliseconds, which would make per-frame video analysis practical.
The benchmark picture is messy, and the article doesn't pretend otherwise. On the AIM System One Index and the Artificial Analysis Decision Index, OpenAI's model ranks eighth and fourteenth, behind Liquid's D1 Ultra and Base Lab's Elephant-1, while TypeSafe's latest Jev sits seventh and third. TypeSafe's models are still the most queried on OpenRouter. Where OpenAI may have an edge is multimodal work, particularly image classification, where Handa says Luna shows better calibration. Even so, the article's bottom line is that the right model depends on the workload, and the most reliable test is running candidates against your own tasks.