OpenAI Decisions API Trades Prose for Fast Typed Verdicts

OpenAI has opened a new interface in public beta — the Decisions API, a dedicated endpoint that returns typed judgments instead of generated text. The launch, announced on October 6, gives developers working with autonomous systems a faster, cheaper way to get structured verdicts from a model — a probability, a choice, or a score — without waiting for prose.

The OpenAI Decisions API runs exclusively on the gpt-6-luna model during the beta period. According to OpenAI's API changelog, the feature first appeared as a limited preview at DevDay 2026 in late September before graduating to public beta. OpenAI says general availability will follow in the coming weeks.

Three question types instead of one long answer

Rather than asking a model to write out reasoning and then extracting an answer from the text, the Decisions API accepts text or image input plus a list of questions, each typed in one of three forms. Predicate questions return the probability that a condition is true. Choice questions select a single option from a fixed list and report confidence scores for each. Score questions rate input against ordered levels, such as severity ratings.

The format matters because a large share of agent work is classification, not composition. Routing a support ticket, flagging a damaged product photo, or deciding which queue a task belongs in are judgment calls that end in a discrete verdict. According to CellCog, images must be supplied as inline base64 data URLs; hosted image links and uploaded file IDs are refused in the current beta.

By skipping token-by-token generation, the endpoint sidesteps what has long been the most expensive step in these pipelines. OpenAI says the API produces answers about ten times faster than asking the same question through its Responses API. That claim, reported by multiple outlets covering the launch, is measured against OpenAI's own Responses API rather than an external competitor, and neither the announcement nor follow-on coverage publishes head-to-head accuracy numbers.

The economics of verdicts at scale

Pricing is input-only: ten cents per million input tokens, with no charge for output tokens or for cache reads and writes. According to AI Weekly, that structure cuts total cost to roughly one-fifth of what standard Luna pricing implies for classification workloads where answers are short.

The trade is deliberate. For a pipeline that processes tens of thousands of classification requests a day, output-token charges on a prose response are pure overhead — the application discards the prose and keeps the label. The OpenAI Decisions API removes that step entirely, which changes the economics of automated moderation and routing. Ground Truth notes the endpoint supports up to 128 inline images per request, and that regional and long-context modifiers can still apply to the base price.

OpenAI is not alone in chasing this market. TypeSafe AI's Jev, a decision model that spread through Vercel's AI Gateway in September, lists input pricing of about four cents per million tokens. According to MarkTechPost's comparison table, a million 500-token classifications would cost around fifty dollars on OpenAI's API versus roughly twenty-one dollars on Jev. Neither vendor has published head-to-head accuracy figures, so the real comparison will have to wait for independent benchmarks.

Why agent builders are paying attention

The Decisions API arrives as agents move from demos into production, where deterministic control matters more than fluency. A typed verdict is a cleaner control primitive than a parsed sentence: it slots directly into thresholds, audit logs, and human-review queues. According to positionmysite.ca's analysis, the API is one narrow call inside a larger routing workflow — intake, redaction, thresholds, the action itself, and the audit trail all remain the builder's responsibility.

Compliance features suggest the OpenAI Decisions API is aimed at enterprise deployments, not just hobbyist projects. The beta ships with Zero Data Retention support and HIPAA eligibility for qualifying customers, plus data residency in the United States and Europe. According to CellCog, Canadian teams in particular need a privacy check first, since OpenAI lists regional processing only for the United States and the European Economic Area. That kind of control logic is becoming a recurring theme in production agent deployments — from meeting agents that now sit inside enterprise workflows to open-source agent frameworks pushing into corporate environments.

The rough edges worth knowing

The beta has quirks that builders should read carefully. One engineer at Towards AI found a rounding trap in the score type: the score is a probability-weighted average across levels, not a level itself, and rounding it can name a level the model assigned zero probability to. In the analysis, more than a third of three-level probability splits on a fine grid showed disagreement between the rounded score and the most likely level. The fix is to read the underlying probability distribution rather than the headline number.

Other limitations are documented in the beta itself. Only one model is supported. The ten-times-faster claim ships without a benchmark or per-call compute disclosure. And the question set is deliberately narrow: this is a classifier endpoint, not a general reasoning API, so anything requiring multi-step deliberation still belongs on the full Responses stack. The incident-prone nature of autonomous systems has also been in the news recently, with reward-hacking episodes pushing at least one lab to restrict agent internet access — another sign that the industry is building guardrails alongside capabilities.

Still, the direction is hard to miss. As AI Weekly's coverage puts it, businesses deploying large numbers of agents need models that can perform routine judgments affordably and consistently. An endpoint built to answer rather than to write is a small but telling sign of where the agent economy is heading: away from chat and toward typed, auditable decisions that machines can act on directly.