Updated 16 Nov 2026: Decisions API GA Oct 22: 400 questions/request (was 200), image instructions, input_json and shared-input caching on GPT-6 Luna.

Key details

  1. Announced September 29, 2026 at OpenAI DevDay 2026.
  2. Available in limited preview with broad release planned in the coming days.
  3. Uses GPT-6 Luna for bounded decision workloads.
  4. Developers define questions with finite pre-defined answers.
  5. Context can include text or images.
  6. OpenAI names classification, request routing and agent next-action selection as example uses.
  7. October 22: Decisions API GA on GPT-6 Luna, with up to 400 questions/request, image instructions, input_json and shared-input prompt caching.

What builders should take away

  1. Benchmark the Decisions API against your existing classifier or structured-output LLM path, not against chat quality.
  2. Measure calibration, latency, cost per accepted decision and escalation rate on labelled production-like data.
  3. Keep deterministic business rules outside the model where the decision does not need probabilistic judgement.
  4. Treat limited-preview schemas and behavior as changeable until OpenAI documents the broad release.

What changed

At DevDay on September 29, 2026, OpenAI introduced Decisions API in limited preview. Developers define questions with finite pre-defined answers, provide context as text or images, and receive answers intended for direct software use. OpenAI names content classification, request routing and selecting an agent's next action as example workloads. The API uses GPT-6 Luna and OpenAI says a broad release is planned in the coming days. UPDATE OCTOBER 22, 2026: OpenAI moved the Decisions API from limited preview to general availability with GPT-6 Luna. The GA endpoint supports up to 400 questions per request (up from 200 in public beta), reusable input caching for prompt-cache discounts, per-question image instructions, and structured input_json payloads. The September 29 limited-preview status is superseded; model calibration and independent reliability evaluation remain open questions.

Why it matters

This is not merely another structured-output option. OpenAI is exposing a separate API surface around the idea that many production AI calls are bounded decisions rather than conversations. That is the same software layer recently targeted by Jev, CLM-8B and GLiNER2.5-Decide. A frontier API provider adopting the pattern makes it easier for builders to compare decision-specific calls against general LLM generation on latency, cost, calibration and operational simplicity.

The contract is finite answers, not generated prose

Developers supply context and a known answer set instead of asking a model to generate and then parse free-form text. OpenAI explicitly positions the interface for classification, routing and agent-action selection, where the application already knows the legal outcomes.

Luna becomes more than a cheap chat model

The first Decisions API implementation focuses GPT-6 Luna on bounded questions. That gives OpenAI a way to reuse its inexpensive model tier behind a task-specific interface while applications consume decisions rather than conversational output.

The category is getting crowded quickly

BTN has already tracked Jev's typed probabilistic API and external Vercel adoption, open CLM-8B, and the much smaller GLiNER2.5-Decide. OpenAI's entry does not make those systems equivalent, but it strengthens the case that decision-specific inference is becoming a distinct production layer rather than a single startup's API design.

The important evidence is still missing

OpenAI has not yet published enough detail in the DevDay recap to compare calibration, latency, per-decision economics or exact response semantics against Jev and other decision models. Limited-preview behavior may also change before broad release.

October 22 GA expands batch size and structured inputs

OpenAI's official API changelog confirms Decisions API GA with gpt-6-luna. Requests can contain up to 400 bounded questions, up from 200 in public beta, and can use shared-input prompt caching, images in individual question instructions and direct input_json. Builders should re-evaluate throughput, cache reuse and schema validation for routing/classification workflows; the release does not itself establish improved accuracy or calibrated probabilities.

What to watch next

  • Whether OpenAI returns probabilities or calibration information alongside answers.
  • Independent comparisons with Jev, CLM-8B and GLiNER2.5-Decide.
  • SDK and gateway support for a first-class decision primitive.
  • Production case studies replacing repeated general-LLM calls with bounded decisions.
  • Independent reliability and calibration measurements for 400-question Decisions API batches.

Still unclear

  • The DevDay recap does not yet provide full API schema, pricing or latency details.
  • OpenAI has not published independent calibration or decision-quality evidence.
  • Limited-preview behavior may change before broad release.
  • It is not yet clear how much of the behavior comes from Luna itself versus a task-specific serving or prompting layer.

Sources

Direct reading behind this dossier.

2 sources
DevDay 2026 Recap
OpenAI primary

Primary announcement for Decisions API, Luna basis, finite-answer contract, supported context and limited-preview availability.

Discussion

Discussion is reader-contributed. Comments are not part of the BTN dossier or its editorial evidence.

0 visible comments

Join the discussion

Keep comments useful and relevant. Reader contributions may be moderated and are not BTN editorial evidence.

Sign in to comment