OpenAI Decisions API goes GA: 400 bounded AI decisions per request
The Decisions API gives AI agents a dedicated way to classify, score and route inputs. The opportunity is faster orchestration, not permission to execute sensitive actions.
OpenAI made the Decisions API generally available on October 22, 2026. Powered by GPT-6 Luna, it accepts up to 400 independent bounded questions per request, alongside text, images and structured input. The current developer guide lists $0.10 per million input tokens on this endpoint, with no separate output or cache-read/write token charges; regional and long-context multipliers may apply.
CLAIM-LEVEL EVIDENCE
What is evidenced and still open
First-party announcements, vendor claims and open questions remain separate.
- Primary source
General availability confirmed
OpenAI states the API became generally available on October 22, 2026.
Source: OpenAI – Build more responsive apps with the Decisions API ↗
- Primary source
400 independent questions
The guide specifies independent batches and requires separate requests for chained decisions.
- Vendor claim
Up to 10x faster
This is OpenAI's comparison to GPT-6 Luna via Responses, not an independent guarantee.
Source: OpenAI – Build more responsive apps with the Decisions API ↗
- Open question
Production errors and real cost
False positives, missed detections and end-to-end spending need workload-specific evaluation.
Vendor releases are not independent effectiveness tests. Methodology →
Why agents need decisions, not always prose
Not every AI action requires a lengthy answer. Support routing, document classification, risk labels and the next step of a customer-service workflow often depend on a tightly constrained choice. OpenAI has now released the Decisions API generally, based on GPT-6 Luna, with an interface designed for predicates, selections and scores. This is a meaningful distinction from a conventional generative response: the model helps estimate a decision inside a predetermined answer space. It does not acquire authority to delete records, execute code or spend money. Those actions belong to the host application's permission system.
A 400-question batch is not a complete agent
The API accepts up to 400 independent questions in a single request, double the previous beta limit. Multiple questions can evaluate shared text, image content or structured data without rewriting all context for each call. There is an essential implementation catch: dependent decisions must run in separate requests. A classifier cannot condition its second step on an answer that has not yet been returned within the same independent batch. Tool orchestration therefore still needs explicit branching, validation, fallbacks and error handling. Output types include predicate probabilities and distributions over choices or scores, not guaranteed factual truths.
Pricing, supported regions and data controls
The developer guide lists $0.10 per million input tokens for gpt-6-luna on /v1/decisions. It currently says there are no additional charges for output, cache reads or cache writes on that endpoint; applicable regional premiums and long-context multipliers remain relevant. Native JSON, images and a GPT-Live integration broaden use cases beyond text-only forms. OpenAI documents regional processing in the United States and Europe, including the EEA and Switzerland. Zero Data Retention and HIPAA use are available only to eligible customers with the required controls and agreements. Regional processing is not itself an independent audit of every tool or downstream storage system.
Why the speed claims need context
OpenAI advertises up to ten times the speed of GPT-6 Luna through the Responses API. It also quotes early deployments from companies including Box, Datadog, Ramp and Sierra, which report their own improvements for different tasks. These are useful examples, not a single controlled independent benchmark. An agent processing ten thousand easy labels may see very different results from one deciding whether a risky shell command can proceed. False negatives and calibration errors can dominate nominal cost savings. The correct metric is cost and latency per accepted, safe outcome rather than model throughput by itself.
A safer way to pilot bounded decisions
Teams evaluating the API should assemble labeled cases from their actual workflow, define acceptable confidence thresholds and measure both incorrect approvals and unnecessary escalations. Route ambiguous outcomes to a reviewer. Then compare a Decisions implementation with a small conventional classifier and a Responses-based approach under the same conditions. Log version, region, prompt changes and downstream effects so results can be reproduced. The wider shift is architectural: fast probabilistic judgements can become a dedicated layer within AI agents, while execution permissions and accountability remain elsewhere.
EDITORIAL ASSESSMENT · NOT INDEPENDENT TESTING
AgentenCode analysis
A dedicated decision layer may make agents substantially more responsive, but measured accuracy, escalation design and authorization boundaries matter more than the headline latency multiplier.
Primary sources
Original announcement: 2026-10-22; editorial publication: 23.10.2026.
Frequently asked questions
Is the Decisions API generally available?
OpenAI announced general availability on October 22, 2026; account, region and processing options remain relevant.
Can 400 questions form a sequential workflow?
No. A single request supports independent questions; dependent choices require separate API calls.
Does it charge for output tokens?
The current Decisions guide lists $0.10 per million input tokens and no separate output or cache-read/write charges for this endpoint, subject to other pricing multipliers.