AI Industry

OpenAI's Decisions API Goes GA at $0.10 per Million Tokens

Oct 23, 2026·5 min read·By Rahish Saifi

Three small stacks of printed expense receipts topped with blank green, amber and coral cards on an oak office desk, as a hand drops one more receipt on the middle stack

On October 22, 2026, OpenAI moved its Decisions API from public beta to general availability, giving every customer a dedicated endpoint for the small, high-frequency judgments that agents and applications make thousands of times a day—routing, classification, risk scoring, and the next step in a voice session. Built on GPT-6 Luna, the API answers bounded questions from text, images, and structured inputs, and OpenAI says it can run up to 10× faster than calling GPT-6 Luna through the Responses API for the same kinds of decisions. The launch lands as rivals push their own “decision” or judgment-focused models; TechCrunch notes OpenAI announced the API at its developer day on September 29, 2026 and opened it to all customers on October 22, 2026.

Key facts

Item	Detail
Status & date	Generally available as of October 22, 2026 (OpenAI product post; API changelog)
Model	gpt-6-luna only—sole supported model on `POST /v1/decisions` (OpenAI developer docs)
Public beta	Decisions API beta launched October 6, 2026 (API changelog)
Speed claim	Up to 10× faster than GPT-6 Luna via Responses API for bounded decisions (OpenAI product post)
Answer types	Predicate (probability 0–1), choice (one of your values), score (probability-weighted average of rubric level indices) (OpenAI product post)
Batch size	Up to 400 questions per request—double the prior beta limit (OpenAI product post)
GA additions	Native GPT-Live integration, prompt caching, US/EU inference residency, native JSON inputs, images in question instructions, more Asia capacity (OpenAI product post)
Pricing	$0.10 per million input tokens; $0.01 per million cached input tokens; no output-token charges (OpenAI product post)
Data	Zero Data Retention for eligible customers; regional processing in US and EU (OpenAI product post)

What the Decisions API is

OpenAI frames Decisions as a way to focus Luna's intelligence on questions with finite answer sets instead of open-ended generation. You define a question, supply allowed answers or rubric levels, and pass context as text, images, or JSON. The API returns a structured result your code can branch on—probability that a condition holds, a single selected option (with probabilities and confidence), or a score along an ordered rubric. That pattern fits agent loops where latency and cost compound: each routing hop, safety check, or “what next?” call adds up if you pay for full chat completions every time.

The mental model is closer to a classifier or policy head than a general assistant. OpenAI's own teams use it internally for categorization, coverage scoring, and faster product interactions, according to the October 22 announcement. For builders, the dedicated /v1/decisions path is the supported integration surface; the developer guide and changelog document the request shape and the single-model constraint.

What changed at general availability

Since the October 6 public beta, OpenAI added several production features ahead of GA. GPT-Live now integrates natively, so voice apps can fire decisions on live transcript events—for example, picking an avatar expression from a fixed palette as the user speaks. Prompt caching lowers cost on repeated context. Inference residency in the United States and European Union addresses data-sovereignty requirements for regulated workloads.

On the input side, developers can pass native JSON, embed images inside question instructions, and pack up to 400 questions into one request. OpenAI also cites continued latency work, including additional capacity in several Asian regions—relevant for global agents that need sub-second decisions close to users.

Pricing and data controls

Decisions is priced aggressively for high-volume micro-judgments: $0.10 per million input tokens, with cached input at $0.01 per million. OpenAI charges only for input tokens; there are no output-token fees, which matches the API's compact structured responses. OpenAI's docs note that cache writes and regional processing are billed separately. Customers who need stricter handling can use Zero Data Retention where eligible, alongside regional processing options documented in OpenAI's pricing and data-control guides.

What early customers report

OpenAI's GA post quotes five companies with specific benchmarks from their own evaluations. Treat these as vendor-reported, workload-specific results, not independent benchmarks—but they illustrate where teams are placing Decisions in production:

Box — Biswaroop Palit, VP, Product Management Platform and Integrations: with Box Extract, Decisions delivered ~5× faster route selection on 30 certificate-of-insurance cases, with the same accuracy as the Responses API.
Datadog — AJ Stuyvenberg, Staff Engineer: scoring risk of agent-generated rshell commands before execution saw 40× faster p95 versus available frontier models, with comparable accuracy, enabling use on the remediation hot path.
Ramp — René Sultan, Research Engineer: AI cost attribution became 4× faster and 30% cheaper versus GPT-6 Luna, without sacrificing accuracy.
Sierra — Ajeet Grewal, Engineer: choosing the next action during parts of its orchestration in live customer conversations runs at under 280 ms p95, about 3× faster and 90% lower cost per call than their previous approach, with no drop in customer outcomes.
Speak — Andrew Hsu, CTO: GPT-Live integration cut time to next action from 1.4s to 0.13s (over 90% lower), improving interactivity for language learners.

Who should use it—and who should wait

Decisions is a strong fit when your product makes repeated, schema-like choices under latency pressure: document routing, command-risk gates, cost attribution tags, conversation orchestration, and real-time voice reactions. If you already batch hundreds of labels per request or share long system prompts across questions, the higher per-request question limit and caching economics matter.

It is a weaker default when you need open-ended reasoning, long-form answers, or model choice. Today only gpt-6-luna is supported; teams standardized on other OpenAI models or multi-vendor routing must branch explicitly. Workloads that require rich tool use or multi-step planning still belong on Responses or Agents APIs, with Decisions as a specialized sidecar.

Caveats worth planning for

First, single-model lock-in on the Decisions endpoint simplifies operations but removes fallback if Luna is wrong for your domain. Second, the headline speed and cost figures come from OpenAI and customer self-evaluations; your prompts, context size, and region may differ. Third, bounded answer types require upfront schema design—predicate, choice, and score are powerful but unforgiving if your taxonomy is ambiguous. Finally, compliance features such as ZDR and EU/US residency apply only where your account and product tier qualify; verify against current data-control docs before promising residency to end users.

Bottom line

OpenAI's Decisions API is generally available as of October 22, 2026, packaging GPT-6 Luna for fast structured judgments at $0.10 per million input tokens (and $0.01 for cached input) with no output-token billing. GA brings GPT-Live hooks, caching, residency, richer inputs, and doubled batch size, while early adopters cite large latency and cost wins on routing, safety, attribution, and voice. For agent builders, it is less a replacement for chat completions than a dedicated lane for the decisions that should never have been full generations in the first place.