OpenAI launches Decisions API: pick answers instead of text, powered by GPT-6 Luna
OpenAI's new Decisions API answers ~150ms decisions from predefined options on a specialized GPT-6 Luna — no prose, just a pick plus a confidence score, in limited preview now.
What happened
At DevDay 2026 on September 29, 2026, OpenAI unveiled the Decisions API, an endpoint built on a specialized version of GPT-6 Luna that does not generate free text. Instead of an open prompt, developers pass a context (text or image) plus a closed list of predefined answers, and the model returns the most likely option with a confidence score — no paragraph to parse afterwards.
Why it matters
OpenAI says the Decisions API returns classified answers in about 150 milliseconds, versus 1.6 seconds for a standard call to GPT-6 Luna — roughly 10x faster. That makes it a fit for high-volume workloads where routing every micro-decision through a general-purpose model wastes both latency and money. The launch also pulls OpenAI into the decision-model category — models that return picks, numbers, and classifications instead of prose — where TypeSafe AI's Jev has been running for two weeks with open-source clones already spawning.
How fast
OpenAI says the Decisions API returns classified answers in about 150 milliseconds, versus 1.6 seconds for a standard call to GPT-6 Luna — roughly 10x faster. That makes it a fit for high-volume workloads where routing every micro-decision through a general-purpose model wastes both latency and money.
What it's for
Typical use cases: classifying content, routing customer requests (e.g. support vs. sales vs. complaints queue), moderation decisions, and letting an AI agent pick its next move from a fixed set of options. Quoting OpenAI Developers on launch day: "Give your app real-time decision-making with Decisions API, powered by GPT-6 Luna. Define questions and possible answers to classify content, route requests, or choose an agent's next action."
Availability
The API entered limited preview at DevDay 2026 on September 29, 2026, with a broader rollout promised in the coming days. OpenAI has not yet published pricing for the preview.
Why now
The launch lands two weeks after TypeSafe AI came out of stealth with Jev, a non-conversational "decision model" that returns numbers (probabilities, scores, classifications) instead of prose — a category that already spawned several open-source clones. OpenAI is arriving third in that month-long race, after Jev and a Stanford/NVIDIA entry, but with the widest distribution.
One caveat worth knowing
Independent reporting notes a real technical distinction: Jev calibrates its confidence scores against actual outcomes, while OpenAI's API returns model self-reported probabilities. Until OpenAI publishes calibration data, treat the confidence score as indicative, not a true probability.
Sources
- https://pasqualepillitteri.it/en/news/19372/openai-decisions-api-jev
- https://www.worthview.com/openai-devday-2026-all-the-biggest-announcements-from-dots-and-gpt-6-1-sol-to-the-agents-api/
- https://www.everydev.ai/p/blog-dot-dot-dot-done-what-actually-shipped-at-devday-2026
Facts checked against the sources above on 2026-09-29; figures (150ms vs 1.6s, limited preview) come from DevDay 2026 launch coverage and the OpenAI Developers announcement.
Verified September 29, 2026.