OpenAI Decisions API: What Fast Typed AI Answers Cost You

OpenAI has moved its Decisions API to general availability, and it is one of the more quietly important developer releases of the autumn. Announced on October 22, 2026, the Decisions API is a dedicated endpoint that asks GPT-6 Luna narrow, bounded questions and gets back typed answers — a probability, a selected option, or a rubric score — instead of free-form prose. For anyone building classifiers, routers, moderation queues or agent orchestration, it turns a slow, verbose model call into something closer to a function that returns a number.

• Availability: generally available since October 22, 2026, after a DevDay preview on September 29 and a public beta on October 6
• Model: GPT-6 Luna is the only supported model, via the POST /v1/decisions endpoint
• Price: $0.10 per million input tokens and no charge for output tokens, according to Unite.AI and Crypto Briefing
• Scale: 1 to 400 questions per request, double the beta limit, with OpenAI claiming up to 10x faster answers than asking Luna through the Responses API

What Is the OpenAI Decisions API?

Most model calls return paragraphs an application must parse; the Decisions API flips that around. A developer sends shared input — a text string, an image, structured JSON, or an entire conversation history including tool calls — and attaches a list of named questions. The response is an array of answers that echo those names, so the results drop straight into application logic.

There are three question types. A predicate question returns a probability between 0 and 1 that a condition is true, such as whether a product photo shows visible damage. A choice question picks one of the developer's predefined options and returns a probability for each option plus a confidence value. A score question rates the input against two to ten ordered rubric levels and returns a probability-weighted average, which means a score can sensibly land between two levels.

That design is a natural fit for the "decision model" category we have been tracking, from Microsoft-Decision-1 on Foundry to the calibrated Jev decision model. OpenAI's version brings the pattern to one of the most widely used developer platforms.

How Much Does the Decisions API Cost?

Both independent reports put the headline price at $0.10 per million input tokens, with no output-token charge. Because a decision response is tiny by construction, the cost of a call is driven almost entirely by the context you send in, which makes budgeting unusually predictable. Unite.AI reports cached input at $0.01 per million tokens, with regional processing premiums and long-context multipliers still applying; Crypto Briefing describes cache reads and writes as free. We would treat the input-only billing as settled and check OpenAI's pricing page for the cache rate before modelling a large workload.

For comparison, OpenAI priced GPT-6 Luna at $0.10 input and $0.50 output per million tokens when it launched in September, as covered in our GPT-6 Sol vs Luna guide. Removing the output charge is the meaningful change for high-volume classification.

What Changed Between Beta and General Availability?

The beta period produced real upgrades. According to Unite.AI's summary of the announcement and documentation, OpenAI added:

• Native GPT-Live integration, so a voice app can run decisions as transcripts arrive
• Prompt caching for reused shared input prefixes of at least 1,024 tokens
• Native JSON inputs and images inside question instructions
• US and EU inference residency, Zero Data Retention and HIPAA eligibility for eligible customers
• Organization usage reporting and Private Link support

Files, audio and item references are not supported, and decisions that depend on an earlier answer still need separate requests.

Early Results From Customers

OpenAI's launch quotes, as reported by Unite.AI, show why latency-sensitive teams care. Datadog says it scores the risk of agent-generated shell commands with 40x faster p95 latency than available frontier models at comparable accuracy. Sierra reports under 280 milliseconds p95 for next-action choices in live customer conversations, roughly 3x faster and 90% cheaper per call than its previous approach. Speak says time to next action in GPT-Live fell from 1.4 seconds to 0.13 seconds, and Box reports about 5x faster route selection at the same accuracy. These are vendor-supplied customer figures, not independent benchmarks.

Why It Matters for AI Agents

Agents spend much of their time making small calls: which tool next, whether to escalate, whether an output passes a check. A fast, cheap, typed decision endpoint lets builders put a probability on each of those steps and route low-confidence cases to humans, which is exactly how more reliable AI agents get built. The Decisions Playground lets developers test questions on their own text or images before writing code.

Follow the rest of the model and tooling race in our AI coverage.

Sources: OpenAI — Build more responsive apps with the Decisions API — October 22, 2026; Unite.AI — October 22, 2026; Crypto Briefing — October 22, 2026.