OpenAI launches Decisions API with GPT-6 Luna, 10x faster classification and $0.10 per million input tokens

6 minutes ago

4 min read

OpenAI announced the general availability of Decisions API on October 22, 2026, following an October 6 beta. Powered by GPT-6 Luna, the endpoint returns typed decisions for classification, selection and scoring without generating conventional prose.

OpenAI reports up to 10x faster performance than GPT-6 Luna through Responses API for supported bounded-answer workloads. The service charges $0.10 per million uncached input tokens, $0.01 per million cached input tokens and no output-token fees.

The release expands the maximum to 400 independent questions per request and introduces native JSON inputs, caching improvements, GPT-Live integration, regional processing and enterprise infrastructure options.

··········

HOW DECISIONS API DIFFERS FROM RESPONSES API.

The specialized POST /v1/decisions endpoint accepts a model, shared input and an array of typed questions. It is designed for machine-readable answers rather than conversational generation.

GPT-6 Luna is the only supported model at launch. General-purpose Responses API remains appropriate for open-ended explanations, complex tool workflows and document generation.

........

Capability	Decisions API	Responses API
Purpose	Classification and scoring	General-purpose generation
Output	Typed decision objects	Generated responses
Model	GPT-6 Luna	Supported Responses models
Batching	Up to 400 questions	No equivalent question array

........

The provider-reported tenfold speed advantage is not a guaranteed end-to-end improvement. Network delays, input length, batch size and downstream services affect application latency.

··········

PREDICATE, CHOICE AND SCORE QUESTIONS.

Predicate questions estimate whether a condition holds. Choice questions select a category from specified options. Score questions evaluate an input against a rubric with distinct levels.

Choice and score answers include probability distributions and confidence information. These are model estimates, not verified accuracy rates. Individual questions may receive refusals; production code must validate each answer.

........

Type	Output	Use case
Predicate	Probability of condition	Sensitive-data detection
Choice	Category and probabilities	Support routing
Score	Rubric level and probabilities	Urgency prioritization

........

For support routing, a predicate might detect a replacement request, a choice could select the responsible team and a score could indicate escalation urgency.

··········

REQUEST STRUCTURE, JSON AND MULTIMODAL INPUT.

Developers send POST /v1/decisions with model gpt-6-luna, shared input and a questions array. A predicate named requires_replacement could ask whether the message 'My package arrived damaged and I need a replacement' requests a replacement.

The response includes answers, model details and usage. Applications should validate names, types and probabilities before acting.

Native input_json parts support structured records without converting every field into prose. Images may be provided using HTTPS URLs or inline base64 data URLs; file IDs are not supported for image inputs. Up to 128 images may be included subject to applicable limits.

··········

BATCHING UP TO 400 INDEPENDENT QUESTIONS.

The 400-question limit doubles the beta ceiling of 200. Multiple questions can evaluate shared evidence such as an invoice, product image or customer request.

Questions within the same request are independent and cannot rely on earlier answers from that request. Sequential workflows require separate requests and application orchestration.

Large batches can still consume substantial input processing. Teams should benchmark latency and token usage rather than assuming the maximum batch size is optimal.

··········

PRICING AND COST PER DECISION.

The API bills input processing at the published Decisions-specific rates, with no output-token charges. Cache writes and regional or long-context processing may introduce additional charges.

........

Component	Rate	Condition
Uncached input	$0.10 per 1M tokens	Standard input
Cached reads	$0.01 per 1M tokens	Eligible reused context
Output	No charge	Decisions pricing
Cache writes	Separately charged	By configuration

........

One million decisions using an average of 500 uncached input tokens each consume 500 million tokens, costing approximately $50 at the base rate. At 2,000 tokens each, the same volume would cost approximately $200 before additional charges.

These calculations illustrate token pricing rather than measured production expenditure. Shared input, caching and batching may change the economics.

··········

PROMPT CACHING AND SHARED CONTEXT.

Prompt caching can discount repeated shared prefixes. The documented minimum reusable prefix is 1,024 tokens; question-specific content does not count toward the shared prefix.

The prompt_cache_key can partition reuse, while explicit breakpoints let developers control eligible cache boundaries. In explicit mode, unmarked boundaries do not receive caching.

Usage details include cached_tokens and cache_write_tokens, which are already components of total input-token accounting and should not be double-counted.

OpenAI also describes 30-minute and 12-hour cache-write options; pricing depends on the applicable caching configuration.

··········

GPT-LIVE AND REAL-TIME APPLICATION CONTROL.

GPT-Live integration allows voice applications to invoke bounded decisions while handling speech interactions, including when transcripts become available.

A voice agent can classify requests and select a workflow without generating a separate user-facing explanation for every internal control step.

Classification probabilities do not establish user identity or authorization. Payments, account changes and sensitive disclosures require separate security controls and confirmation where appropriate.

··········

REGIONAL PROCESSING AND ENTERPRISE GOVERNANCE.

OpenAI describes processing options in the United States and Europe, including the EEA and Switzerland, alongside data residency for supported configurations.

Eligible customers may access Zero Data Retention and supported HIPAA-related configurations subject to contractual and technical requirements. Private Link is also among the enterprise connectivity options.

Organization administrators can query GET /v1/organization/usage/decisions with appropriate api.usage.read permissions. num_model_requests counts requests, not the individual questions inside them.

··········

LIMITATIONS AND PRODUCTION EVALUATION.

Decisions API does not replace open-ended generation or arbitrary multi-step agent execution. It supports GPT-6 Luna, typed answers and independent questions.

Probability scores must be evaluated for calibration, precision, recall and false-positive and false-negative rates on representative labeled datasets. A reported probability of 0.9 is not a guarantee of 90% accuracy.

Developers should handle refusals, missing answers, timeouts and API failures. High-impact actions require independent authorization, oversight and safeguards.

··········

IMPLICATIONS FOR AI APPLICATION ARCHITECTURE.

Many production AI systems need frequent routing, filtering and scoring operations around fewer expensive generation tasks. Decisions API separates those bounded operations into a dedicated, low-cost inference interface.

Its commercial usefulness will depend on whether teams can convert existing workloads into clearly specified questions, preserve decision quality and measure meaningful reductions in end-to-end latency and cost.

··········

·····

DATA STUDIOS

·····

datastudios.org