The Decisions API, announced at OpenAI DevDay on 29 September 2026, lets developers ask a question with a fixed set of possible answers and get back a choice quickly. As of 30 September 2026 it is in limited preview, with pricing not yet disclosed. For growth teams, it looks best suited to routing, tagging and triage jobs where you do not need prose.
Details below are from OpenAI's recap and press coverage; the API was in limited preview when I checked.
Key Takeaways
- OpenAI describes it as focusing Luna's intelligence on user-defined questions with finite, pre-defined answers.
- Press reports cite about 150 milliseconds per decision versus roughly 1.6 seconds for GPT-6 Luna through the regular API (OpenAI claim, per The New Stack).
- It returns ranked answers with confidence scores rather than free text.
- Pricing, maximum answer count and fine-tuning support were undisclosed in coverage I read.
- Good candidates: lead routing, ticket and content tagging, intent classification. Bad candidates: anything needing writing or nuanced judgment.
What OpenAI said
OpenAI's DevDay recap says the Decisions API "enables real-time decision-making by focusing Luna's intelligence on a specific set of user-defined questions with finite pre-defined answers." It is "available in limited preview today with a broad release planned in the coming days."
The Decoder describes it as built on a specialized version of GPT-6 Luna and says developers pass in context as text or images to classify content, route requests or decide an agent's next step. The New Stack adds the confidence-score detail and the 150 ms versus 1.6 s comparison, and lists pricing, answer limits and custom fine-tuning as unknowns.
In plain language
You give it a question, for example "Which team should handle this message?", a list of allowed answers, and some context. It replies with the best answer or a ranking, plus how confident it is. It does not write an email or explain itself at length.
The category sits between two things you may already use: prompting a chat model to pick a label, and training a custom classifier. The New Stack notes that decision-style models accept new labels without retraining while giving reliable confidence scores.
Growth use cases that fit
These are my suggestions based on how growth teams work, not OpenAI-published examples. Test each on your own data.
1. Lead routing
Classify inbound form messages into sales, support, partnership, spam. Confidence thresholds let you auto-route high-confidence items and send the rest to a human.
2. Content and comment triage
Tag social comments and DMs as question, complaint, buying signal or noise, so the community manager sees the important ones first. I would keep a human reply step; classification is not response.
3. Search-intent labelling
Label a keyword list as informational, commercial, navigational or transactional at scale. Useful for a content audit where you have 5,000 queries and no patience.
4. Content inventory tagging
Tag existing URLs by funnel stage, topic cluster and update priority, using the page text as context.
5. Email and survey categorisation
Sort free-text survey answers into themes so you can count them, then read a sample from each theme yourself.
6. Agent decision points
Inside a larger automation, a fast decision step can choose the next action, such as "needs refund review" or "safe to auto-reply," before a heavier model does the work.
Where it does not fit
- Writing copy, briefs or replies. It does not generate prose.
- Open-ended research questions with no fixed answers.
- High-stakes decisions such as credit, hiring or legal outcomes. Do not automate those on a classifier score.
- Anything where your labels are ambiguous. If two humans disagree, the model will wobble too.
How it compares with what you already have
| Option | Strength | Weakness |
|---|---|---|
| Rules and keyword filters | Cheap, predictable | Brittle on wording |
| Prompting a chat model to pick a label | Flexible, no training | Slower, output needs parsing |
| Custom-trained classifier | Tuned to your data | Needs labelled data and upkeep |
| Decisions API (preview) | Fast, new labels, confidence scores | Pricing and limits unknown; preview status |
The New Stack also connects the launch to a competing product from TypeSafe called Jev. I have not tested either, so I cannot rank them.
A test plan for when access opens
- Write the question and labels. Keep to three to eight labels with clear definitions and an "other" bucket.
- Build a gold set. Hand-label 200 real items. This is the dull work that makes everything else valid.
- Run and compare. Measure accuracy per label, not just overall; a rare label like "buying signal" can fail while the average looks fine.
- Set a confidence cut-off. Choose a threshold above which you auto-route, below which a human reviews. Check what share of items falls in each band.
- Monitor drift. Re-score a fresh sample monthly. Customer language changes.
- Get the cost per thousand decisions once OpenAI publishes pricing, and compare it with a cheap chat-model prompt.
Latency: why 150 milliseconds matters and when it does not
At 150 ms a decision can sit inside a live interaction, for example choosing which help article to surface as someone types. At 1.6 seconds, users notice.
For batch jobs, such as tagging a backlog overnight, speed barely matters and cost dominates. Do not pay for speed you will not use.
Data and privacy notes
You will be sending customer messages to an external API. Check what your privacy policy says, strip personal data where you can, and read OpenAI's data-use terms for API traffic before piloting. For Indian businesses, also consider obligations under the Digital Personal Data Protection Act; I am not a lawyer, so confirm with one.
Rollout advice for small teams
If you run a small growth team, resist building a big system on a preview API. Start with one narrow job, such as sorting inbound form messages, and keep your existing rules running in parallel for a few weeks. Compare where they disagree; those disagreements are where you learn the most. Write down the label definitions in a shared document so a new teammate would label the same way. Finally, log every automated decision with its confidence score, so when something goes wrong you can trace which items were auto-routed and why, and adjust the threshold instead of guessing.
What I could not verify
- Pricing, rate limits, maximum number of answer options, and fine-tuning support: undisclosed in the sources I read.
- Real-world accuracy. All speed figures are OpenAI's claims as reported.
- Whether broad release has happened; it was "in the coming days" as of 29 September.
I would not commit any budget or roadmap slot until the pricing page exists.
FAQ
What is the OpenAI Decisions API?
It is an API announced at DevDay 2026 that focuses Luna's intelligence on user-defined questions with a finite set of answers. It returns a choice rather than free-form text, in limited preview as of 30 September 2026.
How fast is it?
The New Stack reports OpenAI's claim of about 150 ms per decision, versus roughly 1.6 seconds for GPT-6 Luna through the regular API. These are vendor figures and I have not measured them.
How much does the Decisions API cost?
Pricing had not been disclosed in the coverage I read. Wait for OpenAI's published rates before planning a budget.
Can it replace a custom classifier?
Possibly for many tasks, since reports say new labels do not need retraining. If you need the tightest accuracy on a narrow domain and have labelled data, a custom model may still win. Test both on your gold set.
What is it good for in marketing?
Routing leads, tagging comments and survey answers, labelling search intent and tagging content inventory. These are my suggested uses, not OpenAI-published examples.
Can it write emails or captions?
No, it is designed to pick from predefined answers. Use a chat model for writing and a decision step for routing.
Is it generally available?
OpenAI said it was in limited preview with broad release planned in the coming days as of 29 September 2026. Check the developer docs for current status.
Should I use it for sensitive decisions?
I would not use it alone for high-stakes calls such as hiring or credit. Use confidence thresholds and keep a human in the loop.
How do I know if it is accurate for my data?
Build a hand-labelled gold set of about 200 items, run the API on it and measure accuracy per label. Repeat monthly on fresh samples.
Talk to Younus
I am Younus Fardeen, an India-based organic growth, SEO and content strategist with 4+ years of marketing experience across edtech and startups. If you want help turning tools like this into workflows that genuinely save time, see my work and reach out through the contact form at younusfardeen.in.