October 31, 2026 · Ahmed · More by Ahmed
What OpenAI’s Decisions API Gained Between Beta and GA
Decisions API went GA with EU/US inference residency and prompt caching added since beta, plus a doubled per-request question cap. Here's what changed.
OpenAI made the Decisions API generally available on October 22, 2026, roughly two and a half weeks after it entered public beta on October 6. Built on gpt-6-luna, the API is scoped narrowly on purpose: instead of free-form generation, it answers questions with bounded formats, which OpenAI says makes decisions “from text, images, and structured inputs up to 10x faster than GPT-6 Luna through the Responses API.” For automation pipelines doing high-volume classification, routing, or scoring, that narrower contract is the point — and the GA release changed what’s actually usable in production with it.
Three Answer Types, One Request Shape
The API returns one of three answer types. A predicate call estimates the probability that a condition is true, returning a value from 0 to 1. A choice call selects one of the developer’s supplied options and returns probabilities and a confidence score alongside it. A score call rates an input against ordered rubric levels and returns the probability-weighted average of those levels. All three are structurally different from the open-ended text a Responses API call would generate — useful specifically because the output is typed and boundable, not because it’s more capable.
What's New Since Beta: Caching and Residency
What changed since beta is where the production case gets stronger. OpenAI added “prompt caching for reused context” and “inference residency in the US and EU.” Caching matters at volume: a pipeline that reuses the same system instructions or rubric across thousands of per-item classification calls now pays $0.01 per million cached input tokens instead of the standard $0.10 per million, on every call after the first. Residency matters for anything handling EU customer data — fraud screening, returns triage, or support-ticket routing that touches EU records — because inference can now stay in-region instead of being an automatic non-starter for teams with EU data-handling requirements.
Bigger Batches, and What They Cost
The per-request question cap also doubled, from 200 in public beta to 400 at GA, and developers can now pass native JSON inputs directly via input_json and include images in individual question instructions. Together, that means more small decisions — per line item, per ticket, per record — can be batched into a single request instead of multiplying API calls.
On pricing, OpenAI states it plainly: “Decisions API costs $0.10 per million input tokens. Cached input costs $0.01 per million tokens... You only pay for input tokens: there are no output-token charges.” The GA announcement also references 30-minute and 12-hour cached writes, pointing to a separate pricing guide for the exact rate — that multiplier isn’t stated for the Decisions API itself, and it shouldn’t be assumed to match the 12-hour cache-write premium documented for the separate Responses API surface.
What This Unlocks for Production Pipelines
For a commerce automation pipeline that needs both a compliance story and a cost ceiling, caching and residency are what move the Decisions API from a promising beta to something worth building production classification logic on.