On October 22, OpenAI swung wide open the doors to a new capability interface—the Decisions API was announced as fully available. Built on top of GPT-6Luna, this interface does one very focused thing: it channels Luna's intelligence into questions with "bounded answers", helping applications and agents make the final call in a near real-time rhythm based on text, images, and structured data. According to OpenAI, its response speed is up to 10 times faster than calling GPT-6Luna via the Responses API.
Pricing is the most eye-catching part of this release. The Decisions API charges $0.1 per million tokens for input (approximately 0.67 yuan at current exchange rates), and cache-hit input costs only $0.01 per million tokens (approximately 0.067 yuan). Even more crucial is its billing logic: you pay only for input tokens, while output tokens are free. In other words, the more you ask and the longer the answers, the bill only counts what you "feed in". OpenAI revealed that the entire company is already running the Decisions API in its internal products and business workflows.
Since the public beta phase, this capability has continued to grow new muscles. OpenAI added native GPT-Live integration, Prompt caching for reusing context, and inference data residency support covering the US and the EU; developers can now also pass native JSON input directly, embed images in question instructions, and the number of questions allowed in a single request has been raised to 400—exactly twice the previous limit. Latency performance is also being continuously refined, including expanded service capacity across multiple regions in Asia. As the model learns to make decisions quickly and cheaply within a bounded frame, OpenAI has pushed forward another notch the step for AI from "being able to give advice" to "being able to decide for you".