News · 2026-10-07

OpenAI launches a Decisions API that returns choices instead of prose

OpenAI launched its Decisions interface in public beta on October 6, letting developers request bounded choices, probabilities, and rubric scores instead of generated prose. The launch uses GPT-6 Luna and costs $0.10 per million input tokens, without output-token charges. It offers a simpler control primitive for applications, but its confidence estimates and advertised speed need validation on the decisions customers actually make.

Key facts

• The beta adds a dedicated decisions endpoint backed by GPT-6 Luna at launch.
• It supports text and inline images, with a maximum of 128 inline images per request.
• Published pricing is $0.10 per million input tokens; regional and long-context modifiers can apply.
• Primary source: OpenAI’s Decisions guide, recorded in its October 6 changelog.

Sometimes an application needs an answer from a short menu rather than a paragraph. A support ticket must go to billing or technical help. A proposed tool action must be allowed, rejected, or sent to review. A document must be placed on an ordered quality rubric. Asking a general language model to generate and explain those decisions adds text that the application may discard, while creating extra parsing work.

OpenAI’s new interface makes the decision space explicit. A predicate estimates whether a condition is true. A choice returns a selected developer-supplied value and probabilities over the allowed values. A score returns a distribution over ordered rubric levels and their probability-weighted average index. Choice and score responses include a separate confidence field. These are bounded outputs, not arbitrary text packaged in a structured object.

A useful analogy is replacing an essay exam with a marked ballot. The ballot is easier for software to count, and it ensures the result names an allowed option. It does not ensure that the voter understood the question or applied the right rule. An invalid option is harder to produce, but a valid-looking wrong decision remains entirely possible.

The official changelog dates the release to October 6. OpenAI advertises “10x faster” than its Responses interface. The reviewed guide and changelog do not specify the baseline model, request size, hardware, concurrency, sample count, or latency statistic. That is a documented vendor claim with undisclosed measurement, not a reproducible general speedup. Customers should measure their actual end-to-end decision time.

Pricing is clearer. The pricing and availability section says there is no charge for output tokens, cache reads, or cache writes on Decisions requests, while input billing and specified modifiers remain. That can matter for workloads that previously generated lengthy responses. It does not prove a large saving over a tightly constrained call that already emitted only one short answer.

The endpoint is deliberately narrower than Responses. It accepts shared text or user messages containing text and inline images. It does not support non-user roles, tool-call results, files, audio, or item references. Multiple independent questions can share the same evidence in one request; a question depending on an earlier answer requires another call. The response gives answers in question order and does not generate an explanatory rationale.

That absence moves responsibility toward application design. The caller must specify options that cover the possible outcomes, distinguish overlapping classes, and provide an escalation route. A model cannot choose a missing option. If uncertainty or missing information is a meaningful state, the application has to handle it. The existing selective-prediction lesson explains why declining or deferring can be part of a correct decision policy.

Probabilities also need local testing. A reported confidence of 0.9 is not a published guarantee of 90% correctness in a new customer’s prompts. OpenAI advises using labeled application examples and choosing thresholds around the costs of false positives and false negatives. Calibration is the difference between a convincing number and a probability that can support operational decisions.

Today’s research supplies a warning without establishing a bug in OpenAI’s system. In Labels Override Definitions, researchers test open decision models and find that option names can overpower the written rules. Neutral labels improve one policy suite by about 15 percentage points, while some corrupted decisions retain misleading confidence. The study does not test proprietary OpenAI Decisions or Jev. Its practical contribution is an audit idea: change labels while preserving definitions and check whether the intended policy still controls the output.

There is also a real local alternative in the conversation. Strands’ October 1 Decider announcement describes a small model with a pointer-style readout rather than generated prose. The two releases support a wider category shift toward models that choose. They do not establish that OpenAI uses the same architecture or that one route is universally better.

The Hacker News discussion includes benchmark requests, cost comparisons, and narrow personal anecdotes. None settles production reliability. The strongest counterargument to the launch’s simplicity is that a clean software interface can conceal a bad policy judgment. Decisions is useful when the task really is bounded; the surrounding system must still prove that its bounds and error handling match the job.

Primary source, verified: read the paper →

Ground Truth. (2026, October 7). OpenAI launches a Decisions API that returns choices instead of prose. Ground Truth. https://groundtruth.day/news/openai-decisions-api-beta-input-only-pricing.html

@misc{groundtruth:openai-decisions-api-beta-input-only-pricing,
  title  = {OpenAI launches a Decisions API that returns choices instead of prose},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/openai-decisions-api-beta-input-only-pricing.html}
}

Topics: tools · decision-models · openai · apis · classification