OpenAI’s Decisions API returns typed answers; the company says it is 10x faster than Responses

What did OpenAI release?

The 6 October changelog says OpenAI released the Decisions API in beta with gpt-6-luna, and that it can “turn text and images into typed answers 10x faster than the Responses API.” That speed claim is OpenAI’s. The dedicated guide repeats the comparison as “about 10x faster,” names a public beta, and says general availability is expected in the coming weeks.

The endpoint is POST /v1/decisions. The only model currently available is gpt-6-luna. A playground is live at platform.openai.com/decisions. Official SDK examples require Python 3.26.0, JavaScript 7.30.0, Go 3.73.0, Ruby 0.101.0 or Java 4.78.0, or later.

The same changelog day simplified API usage tiers from five to three: Build, Launch and Grow, with automatic upgrades as total credit purchases reach each tier minimum. On 5 October OpenAI also added an in-product HIPAA Business Associate Agreement flow for eligible organizations. Those are separate platform changes, not part of Decisions itself.

How do Decisions requests work?

A request has three parts: the model, a shared input, and a questions array. The input can be a text string or user messages that mix text with images. Each question needs a unique name so the answers array can echo it back.

There are three question types. A predicate checks a condition and returns a probability from 0 to 1. A choice selects one of the values you supply, with a probability for each option and a separate confidence field. A score rates the input against ordered levels and returns the probability-weighted average of their indices, so the number can fall between levels. That is closer to a constrained classifier than to free-form structured output.

The guide’s examples are concrete. A predicate can ask whether a product photo shows a crack, tear or dent. A choice can route “I was charged twice for my order” to billing, technical, shipping or other. A score can place “Export fails in Safari but works in Chrome” on a three-level severity rubric. Several independent questions can share one input; questions that depend on an earlier answer need a second request.

How should teams interpret the answers?

A predicate probability is the model’s estimate that the condition is true, not a measured frequency. Choice and score answers add a confidence field beside the distribution. OpenAI tells developers to set thresholds from labeled examples in their own application, and to weigh the cost of false positives against false negatives.

The illustrative responses in the guide are examples, not benchmarks. A visible-damage probability of 0.92, a billing choice at 0.95, or a severity score of 1.1 show the shape of the JSON, not how the model will behave on your photos or tickets.

Answers can also be refusals. Any production path needs a branch for that case, the same way an agent workflow needs an explicit stop when a tool call is declined. Our notes on model routing are relevant if Decisions sits beside a generation model rather than replacing it.

How do you access it, and what does it cost?

Images must be inline base64 data URLs. Hosted HTTP or HTTPS image URLs and file_id inputs are not supported. Combine input_text and input_image parts in a user message when the question depends on both. Token accounting still matters for cost; see our API cost guide for the usual input-token arithmetic.

Decisions pricing is not the same as ordinary gpt-6-luna chat pricing. The Decisions guide says input costs $0.10 per 1 million tokens, and that you pay only for input tokens: there are no cache-read, cache-write or output-token charges. Regional processing premiums and long-context input multipliers still apply. Other gpt-6-luna endpoints keep the standard rates listed on the pricing page, including $0.50 output per million tokens on short-context Standard requests.

The guide says Decisions supports Zero Data Retention and HIPAA use for eligible customers, and data residency plus regional processing in the United States and Europe (EEA and Switzerland). Eligibility, agreements and limits sit in OpenAI’s data-controls documentation, not in this article.

When should you not use Decisions?

OpenAI is explicit: use Decisions when the application needs one of the three answer types. Use Structured Outputs with the Responses API when you need an object that follows your JSON schema, such as extracted fields or a written explanation. Use function calling when the model must request a tool call with arguments. That split matters for AI agents, which still need tools, memory and approval, not only a typed verdict.

Decisions will not write a reply, cite a source, or call your APIs. It will not replace an evaluation set. A routing rule that looks clean in the playground can still mis-file complaints that were never in the labeled sample.

What remains uncertain?

The 10x claim is a vendor comparison with Responses, not an independent latency study. This publication has not timed the endpoint. Secondary reports that mention sub-100-millisecond answers were not in the OpenAI pages we checked, so they are not repeated here.

The beta has one model, no hosted image URLs, and a GA date described only as “the coming weeks.” Anyone putting Decisions on a live path should keep a fallback classifier or human queue until the generally available contract, rate limits and model list are documented.