OpenAI Makes the Decisions API Generally Available With Input-Only Pricing

The Decisions API is generally available now

OpenAI announced the Decisions API’s general availability on October 22, 2026. The production route is POST /v1/decisions, and the current documentation names gpt-6-luna as the only supported model. The endpoint is meant for focused application decisions rather than free-form generation: classify content, route a request, prioritize work or score evidence against a supplied rubric.

Since the public beta, OpenAI has added prompt caching, direct JSON input, images inside question instructions, hosted image URLs, conversation history, organization usage reporting, native GPT-Live integration and Private Link. The per-request ceiling increased from 200 to 400 questions. These are documented service capabilities, not evidence that a particular classifier or routing policy is accurate enough for production.

Three answer types define the API boundary

A predicate returns an estimated probability from zero to one that a condition is true. A choice selects from caller-supplied, unordered options and returns their probabilities. A score maps ordered rubric levels to numeric indices and returns a probability-weighted average plus confidence and the per-level distribution. The application—not OpenAI—decides what threshold triggers review, routing or action.

OpenAI advises using Decisions only when one of those bounded answer shapes fits. Structured Outputs through the Responses API remains the documented path for generating a custom JSON object or explanation, and function calling remains the path for asking a model to propose a tool call. Decisions can inspect prior tool calls and outputs as evidence, but it does not execute the resulting action.

Text, images, JSON and conversations are supported, with gaps

Shared input can include text, images and JSON in user messages, plus ordered conversation messages, function calls, custom tool calls and matching outputs. Developer and system messages accept text only and system messages are treated as developer messages. Files, audio and item references are not supported by the Decisions input contract.

Question-specific images let a caller compare separate visual references against the same evidence. Hosted HTTP or HTTPS images and inline base64 data URLs are supported. Those conveniences expand the external-content and data-governance surface: validate URLs, bound payload size, classify the data and keep authorization outside a model probability.

The endpoint charges input tokens and not output tokens

For /v1/decisions with gpt-6-luna, OpenAI lists $0.10 per million uncached input tokens and no output-token fee. Standard Luna cache rates apply: $0.01 per million cached input tokens and $0.125 per million cache-write tokens. Reusable shared prefixes must contain at least 1,024 tokens; question content sits outside that reusable prefix.

Long-context and regional multipliers still apply. The Luna model reference says prompts above 272,000 input tokens double input and cache rates, while regional processing adds 10% where available. Do not apply the model page’s ordinary $0.50-per-million output price to Decisions, and do not assume “no output charge” means no total cost: repeated evidence, cache writes, regions, retries and review work remain billable or operational costs.

The 10× latency statement is a vendor claim

OpenAI says Decisions returns typed answers about 10 times faster than the Responses API. This is a vendor claim. The reviewed guide does not publish the comparison workload, regions, request sizes, processing tier, sample count, p50 or p95 latency, error rate, answer agreement or raw traces. AccessAllGPT did not run the endpoint, so the statement should not be treated as an independent performance result.

Measure the configured system on production-shaped evidence. Compare end-to-end p50, p95 and p99 latency; timeout and retry rates; token counts; cost; false-positive and false-negative rates; calibration; abstentions or refusals; reviewer effort; and downstream outcomes. A fast answer with a poorly chosen threshold can make the application less responsive to reality, even if the API responds quickly.

Data controls are qualified, not automatic

OpenAI says Decisions supports Zero Data Retention and HIPAA use for eligible customers, with data residency in supported API regions and regional processing in the United States and Europe. Those are eligibility-dependent service options, not a blanket property of every account or request. The broader data-controls guide says default abuse-monitoring logs may contain prompts, responses and derived metadata and can be retained for up to 30 days.

Obtain account-level evidence for the chosen project, region and endpoint before sending regulated or sensitive data. Confirm ZDR or Modified Abuse Monitoring approval, applicable agreements, Private Link configuration, residency, image handling, logs and subprocessors. A product-page statement does not replace a verified control state and signed contractual scope.

Calibrate first, then automate a bounded action

Start with one observable question whose false-positive and false-negative costs are known. Freeze labeled development and holdout cases, define ambiguous and adversarial examples, and compare the returned distributions with human labels. Choose separate thresholds for automatic action, human review and no action; do not convert a probability directly into authority.

Adopt the API for a narrow route only when local calibration, latency, cost, data and failure gates pass. Constrain it behind review when mistakes are recoverable but consequential. Wait when required account controls, stable limits or workload evidence are missing. Reject the path when the decision cannot be expressed as a predicate, fixed choice or ordered score, or when no safe fallback exists.

Disclosures

AccessAllGPT received no OpenAI account, API access, credits, benchmark data, private briefing, review, payment or compensation for this article. OpenAI did not sponsor, review or endorse it. AccessAllGPT did not test, score, rank or deploy the Decisions API or gpt-6-luna. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with OpenAI. Publication-wide relationships are listed on the disclosures page.