What changed
OpenAI introduced the Decisions API in beta on 6 October. It uses GPT-6 Luna to turn text or images into typed answers, and OpenAI says it is ten times faster than the Responses API. That speed figure is a vendor claim, not an independent benchmark.
On 9 October, Vercel added Microsoft Decision-1 and Liquid AI's d1 to AI Gateway. Both return structured choices, scores or boolean answers with probabilities; Liquid d1 also accepts images. They are available through Vercel's AI SDK decision API and compatible HTTP interfaces.
A single request can evaluate several questions against shared input. Vercel also records decision calls in Gateway logs and budgets alongside other model requests. This gives teams a distinct, observable path for classification and scoring rather than treating every control step as open-ended generation.
Our perspective
This makes a useful architectural boundary more explicit. A generative model can create or repair an application, while a narrower decision call can classify an error, score a candidate result or choose the next workflow step without producing prose that another system must parse.
The opportunity is lower latency and more observable control logic, but probability output should not be assumed to be calibrated. Teams should test decisions on representative data, set thresholds for automatic action and send uncertain cases to a stronger model or a person.
Start with a reversible step, such as tagging a build failure or choosing a repair queue. Log the input, answer, confidence, eventual outcome and any override. That evidence shows whether a dedicated decision endpoint improves the workflow or merely moves prompt complexity into a new interface.
Go to the source
Primary reporting behind this brief. Our perspective is editorial analysis.
- OpenAI API changelog — Decisions API beta 6 Oct 2026
- Vercel — Microsoft Decision-1 on AI Gateway 9 Oct 2026
- Vercel — Liquid AI d1 on AI Gateway 9 Oct 2026