OpenAI has moved its Decisions API from public beta to general availability, offering a low-cost endpoint for fast bounded answers such as classifications, routing choices and rubric scores. The October 22 release pairs GPT-6 Luna with text, images and JSON inputs, but independent benchmarks place the offering behind several rivals.

What changed on October 22

OpenAI’s dated product announcement says the Decisions API became generally available on October 22, 2026. That is a meaningful status change from the public beta announced on October 6 in the company’s developer community. The endpoint is available through a dedicated Decisions API rather than as another response mode that developers must assemble themselves. OpenAI’s announcement says it is built with GPT-6 Luna and is intended for questions with bounded answers.

A conventional language-model call may generate a paragraph, a tool call or a custom JSON object. Decisions instead asks for the question and permitted answer space, producing a control signal to classify a document, choose a queue, retry or flag an action. OpenAI’s documentation directs developers to Structured Outputs for a custom JSON schema and function calling when the model must request a tool.

The company says Decisions can make near-real-time choices up to 10 times faster than GPT-6 Luna through the Responses API. That is a company comparison, not an independently reproduced universal result. The baseline is Luna through Responses, not all model tasks.

How the endpoint works

Three typed answer modes

The developer documentation describes three question types. A predicate returns a probability from 0 to 1 for whether a condition is true. A choice selects one of the developer’s predefined values and returns probabilities over the options. A score evaluates an input against ordered levels and produces a probability-weighted average of their indices. These formats are deliberately narrower than an open-ended answer, which is the point: an application can make the next step deterministic in its own code.

Requests use POST /v1/decisions and currently support gpt-6-luna as the model. Shared evidence can be text or an ordered conversation containing messages, tool calls and tool outputs. The endpoint accepts native JSON through an input_json content part and can inspect images supplied as inline base64 data URLs or publicly accessible HTTP or HTTPS URLs. Files, audio and item references are not supported, according to the documentation, so “multimodal” here should not be read as support for every input modality.

What general availability adds

OpenAI says GA adds or expands several pieces from the public beta. Developers can ask up to 400 questions per request, twice the earlier 200. The release also includes native GPT-Live integration, prompt caching, question-level image instructions, hosted image inputs, conversation-history evaluation and usage reporting. Regional processing is available in the United States and Europe, with eligibility and agreement requirements.

GPT-Live integration gives the feature a visible use case. Transcript updates can trigger a choice of an avatar’s expression or help a voice agent ask for clarification, route a request or escalate it. A live session receives a typed decision event without a separate round trip assembled by the developer.

OpenAI lists $0.10 per 1 million input tokens for Decisions with Luna and says output tokens are not charged. Cached input is listed at $0.01 per 1 million tokens, while the documentation notes that cache writes, regional-processing premiums and long-context multipliers can apply. Actual cost still depends on input volume and caching behavior.

Speed, trials and the evidence

What OpenAI and customers report

OpenAI’s announcement includes customer statements, not a neutral laboratory evaluation. Box says Box Extract and Decisions selected routes about five times faster on 30 certificates of insurance cases, with the same accuracy as its Responses API evaluation. Datadog reports 40-times-faster p95 latency than available frontier models with comparable accuracy for assessing agent-generated shell-command risk. Ramp says cost attribution was four times faster and 30% cheaper than with Luna.

Sierra reports under 280 milliseconds p95 in part of its live-conversation orchestration, about three times faster and 90% cheaper per call than its previous approach. Speak CTO Andrew Hsu reports that a GPT-Live integration reduced time to the next action from 1.4 seconds to 0.13 seconds, a reduction of more than 90% in the company’s testing. The units and baselines matter: these are customer-specific comparisons, with different workloads, traffic patterns and reference systems. They should not be combined into a single API-wide performance guarantee.

TechCrunch’s independent report adds a direct account from Nikunj Handa, OpenAI’s product lead for APIs. Handa described shrinking response times from one to three seconds to 100, 120 or 180 milliseconds. TechCrunch presents those figures as the product lead’s description, not as a published percentile benchmark. The report also says the public release reached all customers on Thursday and names Box, Datadog, Ramp and Speak among initial trial users.

What remains unsettled

The same TechCrunch account cautions that the market has no settled definition of a good decision model. Speed is only one dimension; developers may also care about calibration, ambiguous-input accuracy, cost, consistency and context handling. A model excellent for fixed routing may be poor for nuanced safety judgments.

Public comparisons cited by TechCrunch underline that uncertainty. The report says the AIM System One Index and Artificial Analysis Decision Index favored Liquid’s D1 Ultra and Base Lab’s Elephant-1. It places the latest TypeSafe Jev model seventh and third on those two indices, while OpenAI’s Decision Model ranked eighth and fourteenth. The rankings are snapshots of named public indexes, not a definitive verdict on every workload, but they directly qualify the announcement’s performance narrative. TechCrunch also reports that TypeSafe’s Jev was the most widely queried decision offering on OpenRouter at the time of publication.

A new decision-model market

OpenAI’s move follows TypeSafe’s Jev, which TechCrunch describes as a “super classifier” that chooses among options and attaches confidence. The context helps explain why OpenAI exposes a separate endpoint: the product sells an operational shape—small, frequent, typed decisions—alongside the model.

That shape is relevant to agent design. An agent may need a language model to explain a plan, but a faster decision layer can select a tool, filter an input, classify an incident or decide whether a human should intervene. The endpoint reduces generated text to parse and makes allowed outcomes visible in application code. It does not remove evaluation: a confident wrong choice can still send an agent down the wrong path.

OpenAI’s positioning leaves room for change. Handa told TechCrunch that the current system is essentially Luna reconfigured for faster, cheaper inference and that the company may use more specific training. GA is therefore a production-availability milestone, not evidence that the architecture or category has stabilized.

What developers should take from it

The sensible reading is narrow but useful. Teams with repeated, clearly specified decisions can test a dedicated interface instead of forcing a general workflow to emit a verdict. Strong candidates have a known answer set, measurable outcomes and a low-confidence fallback. Image triage, document routing, moderation predicates and voice control fit better than open-ended research or complex planning.

Teams should benchmark against their own data. OpenAI’s 10x statement, customer figures and Handa’s millisecond range use different baselines. The independent rankings show that the fastest or cheapest endpoint is not automatically the most accurate or calibrated. A responsible trial should measure class accuracy, calibration, tail latency, cache effects, regional processing, refusals and human-review cost, including ambiguous cases.

General availability makes those experiments easier to operationalize, but it does not turn vendor claims into guarantees. The durable change is the availability of a low-latency, typed decision primitive from a major API provider. Whether it becomes a default layer in agent stacks will depend on workload-level evidence that is still accumulating.

Frequently asked questions

What became generally available?

OpenAI says the Decisions API, powered by GPT-6 Luna, reached general availability on October 22, 2026. It had previously been in public beta. The GA status applies to the dedicated endpoint, not to a claim that every Luna or Responses API task is faster.

What can a Decisions API response contain?

It can return a predicate probability, a choice from predefined values or a score against ordered rubric levels. Developers define the question and answer space, then use the typed result in application logic.

How fast is it?

OpenAI claims up to 10 times the speed of GPT-6 Luna through Responses. Nikunj Handa told TechCrunch that some response times fell from one to three seconds to 100–180 milliseconds. Customer figures and those ranges are reported claims, not a universal service-level promise.

Does it accept images and JSON?

Yes. OpenAI’s announcement and documentation describe text, images and native JSON inputs. The docs specify inline base64 or public image URLs and an input_json content part. They also say files and audio are not supported.

Does it outperform competing decision models?

There is no supported basis for saying so generally. TechCrunch reports that public indexes ranked OpenAI’s offering behind several competitors in the snapshots it cited. Results can vary by task, and the field is still debating how to measure decision-model quality.

References