OpenAI opened its Decisions API to all customers on 22 October. The model picks from a set of options and returns a confidence score in 100–180 milliseconds, instead of the one to three seconds a standard LLM needs. It answers TypeSafe's Jev, which created the category.
Key takeaways
- Decision latency dropped from 1–3 seconds to 100–180 ms
- Underneath runs ChatGPT Luna, retuned for speed, no separate training
- Speak cut the gap between actions from 1.4 s to 0.13 s
- The public version adds voice and visual input
- OpenAI sits 8th and 14th on two public indexes
Not a new model, just a retuned Luna
Nikunj Handa, product lead for OpenAI's APIs, told TechCrunch the team built no separately calibrated decision model. It took ChatGPT Luna and retuned it for speed and cost.
The big unlock with Decisions API has been the speed, shrinking things from one to three seconds to 100, 120, 180 milliseconds.
Nikunj Handa, product lead for OpenAI's APIs.
Announced on 29 September at the developer day, it also spreads inference geographically to cut latency.
A category that did not exist two years ago
Jev introduced "System One" models — super classifiers?super classifiers: models that pick one option from a closed list and report a confidence score, instead of generating text. that pick one option and report their confidence. Developers swap them in for LLMs where throughput beats eloquence. Box, Datadog, Ramp and Speak did exactly that. Andrew Hsu, CTO at Speak, gave the sharpest number: time between actions fell from 1.4 seconds to 0.13 seconds.
Public rankings are less kind. The AIM System One and Artificial Analysis Decision indices place Liquid's D1 Ultra and Base Lab's Elephant-1 higher. OpenAI's edge is multimodality — Handa says Luna calibrates image classification well.
Why it matters
Shifting decision latency and cost by an order of magnitude changes where calling a model is worth it at all. At a few dozen milliseconds, querying one per video frame becomes reasonable rather than once per interaction. The contest is no longer about text quality but about calibration, latency and unit cost. The market is splitting into two model classes: slow and eloquent, or fast and decisive, picked per task rather than per vendor.
What's next
- Handa signals an attempt to reach roughly 60 ms, opening up per-frame video calls
- OpenAI is weighing dedicated training for the decision mode — for now plain Luna suffices
