OpenAI's Decisions API is now generally available

OpenAI’s Decisions API reached general availability on 22 October 2026, three weeks after its first public preview at DevDay on 29 September. Built on GPT-6 Luna and called via POST /v1/decisions, it is a narrowly scoped endpoint designed for one thing: answering bounded questions fast and cheaply, without generating a full text response.

What the Decisions API actually does

Most AI API calls follow the same pattern. You send a prompt, the model generates tokens one by one, and you get back a chunk of text. For many tasks, that generative step is wasted effort. If you want to know whether a support ticket should go to billing or technical support, you do not need a paragraph. You need a choice.

The Decisions API skips the generation step entirely. You pass a block of content (text, JSON, or images) as state, attach a list of questions, and get back typed, probabilistic answers. There are three question types:

Predicate: returns the probability that a condition is true, for yes/no decisions.
Choice: returns the probability for each option you define, plus the most likely one selected.
Score: returns the probability for each level of an ordered rubric and an expected score.

Because the model reads the input and returns a structured answer rather than generating tokens, OpenAI claims up to 10× faster responses than calling GPT-6 Luna through the Responses API. DevDay coverage put a typical decision at around 150 ms, compared with roughly 1.6 seconds for a standard Luna call.

What changed at general availability

The public beta, which opened on 6 October 2026, already covered the core functionality. General availability on 22 October brought several additions that matter in production:

400 questions per request, up from 200 in the beta.
Images in question instructions, so individual questions can reference visual content, not just the shared input state.
Prompt caching for reused input context, with cached tokens billed at $0.01 per million (versus $0.10 per million for uncached input). The cache kicks in once the shared prefix reaches 1,024 tokens.
Native GPT-Live integration, so apps can trigger decisions as input or output transcript deltas arrive during a live voice session, letting the app act while a user is still speaking.
Inference residency in the US and EU, with additional capacity in several regions in Asia.
input_json for passing structured data directly.
Hosted HTTP and HTTPS image inputs, conversation-history evaluation with tool calls, organisation usage reporting via GET /v1/organization/usage/decisions, and Private Link support.

Output tokens are not charged. Regional processing premiums and long-context multipliers still apply, and Zero Data Retention and HIPAA use are available to eligible customers.

What this means for you

The API is aimed squarely at the classification, routing, and scoring calls that applications and agents make repeatedly inside a single workflow. Asking a frontier model to do that work through the Responses API is expensive and slow. This endpoint is designed to make those calls cheap enough to run on every event.

The real-world results shared at the GA announcement make the difference concrete. Speak, the AI language-tutoring company, cut the time between actions from 1.4 seconds to 0.13 seconds using the Decisions API. Datadog used it to score the risk of agent-generated shell commands before execution and saw 40× faster p95 latency than the frontier models they had been using, at comparable accuracy. Sierra put p95 latency under 280 ms, about 3× faster and 90% cheaper per call than its previous approach. Ramp made AI cost attribution 4× faster and 30% cheaper compared with direct GPT-6 Luna calls.

For developers building agents, the GPT-Live integration is worth noting separately. Agents that need to make routing or scoring decisions mid-conversation no longer have to wait for a turn to complete before acting. That removes a meaningful source of latency in voice-driven agentic workflows.

If you are not building agents or high-volume classification systems, the API is probably not relevant to you yet. It is a specialist tool. But if your application makes the same structured judgements over and over, whether that is moderating content, triaging inbound requests, evaluating outputs against a rubric, or picking the next step in a workflow, the combination of speed, typed outputs, and token pricing makes a serious case for routing those calls here instead of to a general-purpose endpoint.

A competitive context worth knowing

OpenAI is not alone in this space. TypeSafe AI’s Jev model, which spread through Vercel’s AI Gateway in September 2026 and turned TypeSafe into a $7.6 billion company, covers similar ground. Liquid AI, AWS, Perplexity, and Cloudflare all released decision models or APIs between 29 September and 1 October 2026. Vercel’s AI Gateway already exposes an OpenAI-compatible /decisions endpoint that can route to any decision model on the gateway, not only OpenAI’s, which gives developers some flexibility about which provider sits underneath.

The Decisions API reached general availability in under four weeks from first preview. The speed of that release says something about how seriously OpenAI is treating this category.

Getting started

The official decisions guide covers the full request shape and response format. The endpoint accepts up to 1.05 million input tokens per request, and GPT-6 Luna’s context window supports up to 922,000 input tokens in general use. Rate limits on the Build tier are 5,000 requests per minute for GPT-6 Luna; the Grow tier allows 30,000.

If you are already using the OpenAI SDK, switching to the Decisions API requires changing the endpoint and restructuring your payload rather than rewriting your integration from scratch.