OpenAI Releases Decisions API Beta With GPT-6 Luna
The new endpoint accepts text and images and returns typed answers, with a vendor-stated speed gain but sparse public detail about methodology, pricing or operating limits.
Edited by Tyronne Panaino
OpenAI released the Decisions API in beta on October 6, pairing the new endpoint with GPT-6 Luna. The official API changelog says it can turn text and images into typed answers and describes it as ten times faster than the Responses API. That speed comparison is OpenAI's claim; the fetched record does not provide a benchmark method or independent reproduction.
The change gives developers a more narrowly described API surface than a general conversational response. Its stated job is to accept multimodal input and return an answer with a type contract. The announcement establishes the beta release and its intended input-output shape, but it does not document enough in the fetched evidence to infer how broadly the service fits production workloads.
What changed in the API platform
The supported facts are compact. OpenAI added a beta endpoint, named GPT-6 Luna as the model available with it and described text and images as accepted inputs. The output is described as typed, which points developers toward structured application decisions rather than open-ended prose, but the changelog excerpt does not define the available types or schema behavior.
That distinction matters because a beta label is not a general-availability commitment. The source does not state a service-level objective, regional footprint, data-residency options, rate limits or an exit date from beta. Teams evaluating the endpoint therefore have a confirmed product surface, not a complete production contract.
The speed claim needs a workload definition
OpenAI compares the Decisions API with the Responses API and says the new route is ten times faster. The changelog does not identify the workload, request size, response type, latency percentile, hardware or measurement window behind that figure. It also does not say that every decision request will see the same difference.
The defensible reading is limited: OpenAI is positioning the specialized endpoint as a faster path for typed decisions. A universal latency improvement, cost advantage or accuracy gain has not been established by the fetched evidence. Those questions require published methodology or reproducible measurements.
What developers still need to verify
Before treating the beta as a production dependency, teams need details the release note does not supply: pricing, quotas, error behavior, supported answer contracts, model-version stability, image limits, safety evaluation and reliability under load. The source also does not compare output quality with GPT-6 Luna through the Responses API.
Those omissions do not negate the release. They define its current evidence boundary. The next useful checkpoints are a stable reference for the endpoint, transparent performance methodology and a general-availability notice with explicit operating terms.
Evidence quality and status
Confirmed. OpenAI's dated changelog establishes the beta, model, input modes, typed-answer framing and vendor speed claim. Internal confidence is medium because the evidence comes from one first-party source and does not independently validate performance or production readiness.
Sources
Update note: Last reviewed 2026-10-07. We will revise this post if OpenAI publishes endpoint specifications, reproducible performance evidence or general-availability terms.
Sources
- OpenAI API changelog — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.