Netlify routes OpenAI's Decisions API through its AI Gateway

Netlify says developers can now call OpenAI's Decisions API, a beta service that returns structured answers instead of prose, through Netlify's AI Gateway without setting up their own API keys. The announcement appeared in Netlify's changelog on October 7, 2026.

What is new. According to Netlify, code running in Netlify Functions can use the standard OpenAI SDK to classify, route and score inputs, with the gateway handling credentials, so no key management is needed. The company describes the Decisions API as answering questions a developer defines about an input. For "choice" questions, Netlify says the answer is always one of the options the developer declared, which means there is no free-form output to parse before the code branches on the result.

The example. Netlify's sample is a Function for an online store that sorts return requests by reason: wrong size, defective, not as described, or changed mind. Given a customer message about a zipper that failed after two wears, the sample response picks "defective" with a probability of 0.95 and a confidence of 0.93. Netlify suggests sending a prepaid label for defective items, offering an exchange for the wrong size, and using the confidence value to pass unclear requests to a support team. These numbers come from a single illustrative sample, not from a measured test of accuracy.

What is confirmed and what is not. The only source for this story is Netlify's own changelog entry, so every detail here is the company's account. Searches found no independent confirmation of Netlify's rollout; other coverage concerns OpenAI's API generally, or other platforms' separate announcements. Netlify states that the API is in beta and currently runs on gpt-6-luna. It does not say how long the beta will last, whether pricing differs from other gateway usage, or how the service performs on messy real-world inputs. Because the page is a short product note, those questions remain open.

Why it matters. Many production uses of language models are small decisions, such as tagging a ticket or choosing a queue, where free text is a liability. A response limited to declared options is easier to wire into ordinary application logic, and removing key setup lowers the effort of trying it. Whether the confidence figures are reliable enough to automate customer-facing decisions is something teams would need to test themselves.

Revised 16 October 2026: Guth re-read the original sources and rewrote this article. The earlier version remains available as revision 1.