TechCrunch is stamped 12:00 PM PDT · October 22, 2026. OpenAI's Decisions API became available to all customers on Thursday. It had been announced at the developer day on September 29. Product lead Nikunj Handa says the team did not train a separate calibrated decision model. They took ChatGPT Luna and reconfigured it to speed up inference and lower the cost. He later says that what gets the job done right now is plain old Luna. The timing sentence shrinks one to three seconds to 100, 120, and 180 milliseconds. The public version also added voice and visual modalities.

[1]
A charcoal disc lying flat, blank face, one muted blue rim
A charcoal disc lying flat, blank face, one muted blue rim., AI-generated illustration, not a news photograph

Customers named in the initial trials include Box, Datadog, Ramp, and Speak. Andrew Hsu, CTO at Speak, says the time between actions fell from 1.4 seconds to 0.13 seconds. The page prints .13. The AIM System One Index and the Artificial Analysis Decision Index tend to favor Liquid's D1 Ultra, released last week, and Base Lab's Elephant-1. The latest Jev model is seventh and third on those two indexes. OpenAI's Decision Model is eighth and fourteenth. TypeSafe is described as a 7.6 billion dollar company. That is background on this page, not a condition of the release.

[1]

The IT Home page is stamped 2026/10/23 7:48:13. It says the Decisions API is built on GPT‑6 Luna, and that the response is up to 10 times faster than calling GPT‑6 Luna through the Responses API. The 10 times is this page's claim. TechCrunch's 100 to 180 milliseconds, and Speak's drop from 1.4 seconds to 0.13 seconds, are the other page's timings. The two pages do not convert one into the other. The names do not match either: one page says ChatGPT Luna and plain old Luna, the other says GPT‑6 Luna. This story does not collapse them into one model name.

[1][2]

The price appears only on the IT Home page. Input is 0.10 dollars per million tokens, cached input is 0.01 dollars per million tokens, and output is free because the API charges only for input. The yuan figure in parentheses is not in this story. One request can carry up to 400 questions, twice the previous cap. The page also says the preview has since added a native GPT‑Live integration, a cache for reused context, and inference data residency in the United States and the European Union. Neither page prints one table that puts price next to latency. This story does not describe the request format.

[1][2]