OpenAI Decisions API Guide: Instant Yes/No Answers at $0.10 per Million Checks

OpenAI just put a superpower on the shelf: the Decisions API is now generally available (announced in OpenAI’s developer community on 2026-10-22 and confirmed as the opening line of the official docs). In one sentence: it’s a way for your software to ask an AI a quick yes/no, pick-one, or 0–100-score question — and get an instant verdict for a fraction of a cent. If a normal AI chat is like hiring a consultant who talks for twenty minutes, the Decisions API is like hiring a referee who glances at the play and blows the whistle.

What “API” and “token” mean here (30-second plain-language version)

• API (application programming interface): a behind-the-scenes door that lets your software talk to OpenAI’s AI automatically. No chat window — your code sends a question, the AI sends back an answer. Copy-pasting the key is the hardest part.
• Token: the billing unit. Roughly, a word is a token or two. When the docs say “$0.10 per 1 million input tokens,” read it as “$0.10 per roughly 750,000 English words you send in.”
• Predicate, choice, score: the three shapes of decisions this API returns. Predicate = true/false (“is this email a complaint?”). Choice = pick one label (“send this ticket to billing, shipping, or returns”). Score = a number (“rate this review’s anger level 0–100”).

Why should you care?

You probably don’t want to build software yourself. But every small business you know — the Shopify store drowning in customer messages, the dental office sorting appointment requests, the marketplace screening listings — pays humans to do exactly what this API does: look at something and decide what it is. Decisions that used to require a $3,000/month VA or a senior support rep can now be automated for pennies a day. What this means for you: you don’t have to be the one using the API — you can be the one selling the result to businesses that will never touch code. The money section below gives you a concrete recipe.

Price and free quota (prices are vendor-stated, not independently verified)

OpenAI’s official docs list these numbers (all marked as vendor claims — OpenAI says them, nobody has independently audited them):

Item	Price (vendor-stated, not independently verified)	What this means for you
Input	$0.10 / 1M tokens	Checking ~750,000 words costs a dime. A small shop’s entire month of customer messages might cost less than one cent to classify.
Cached input	$0.01 / 1M tokens	If you reuse the same instructions (prompt caching), repeat questions get 10× cheaper still. What this means: high-volume, repetitive sorting is nearly free.
Output	$0 — no output token fee	Normal AI APIs charge you for every word the AI writes back. This one doesn’t, because answers are just a verdict. What this means: your cost per decision is fixed and predictable — ideal for pricing a service on top.
Cache writes	billed separately	One extra line item to watch; still small. What this means: don’t cache carelessly — cache the instructions you actually reuse.

For comparison, the closest rival for this exact job, TypeSafe Jev, charges $0.042 per million inputs flat — cheaper per unit, but OpenAI’s offering brings the brand, the ecosystem, and features like GPT-Live voice integration. US and EU customers can also choose where the AI “thinks” (inference residency), and compliance-heavy customers can use Zero Data Retention (OpenAI keeps no copy of your data). What this means for you: if you sell to healthcare, finance, or EU clients, you can honestly say their data handling meets stricter standards.

New since the public beta (announced 9/29 at DevDay): native GPT-Live voice call integration (trigger decisions during a live phone call — e.g., detect a frustrated caller in real time), prompt caching, native JSON input (feed it structured data directly), images in a single request, up to 400 questions per request (was 200 — so one call can now batch twice the work, halving your overhead), organization usage reports, and Private Link. What this means for you: batching 400 checks per call makes bulk jobs like “classify last quarter’s 40,000 support emails” trivially cheap and fast.

And the headline speed claim: OpenAI says it can be up to 10× faster than the same model on the Responses API (vendor claim, not independently verified). The intuition is simple — the AI skips writing a long answer, so there’s almost nothing to wait for. What this means for you: real-time uses like live-call routing actually work; nobody waits on hold for a verdict.

Hands-on tutorial: your first decision in 10 minutes

Step 1 — Create an OpenAI platform account. Go to platform.openai.com and sign up with your email (it’s the same company behind ChatGPT; the platform site is the developer side). This takes about two minutes.

Step 2 — Get your API key. After logging in, open the “API keys” page (under your profile menu), click Create new secret key, give it a name like “decisions-test,” and copy the key immediately — it shows only once. Treat it like a password; anyone with it can spend your money.

Step 3 — Add a little credit. In Billing, add a payment method and a small top-up (e.g., $5). Decisions are so cheap that $5 will last you thousands of test runs.

Step 4 — Send your first decision. Open a terminal (on Mac: search “Terminal”; on Windows: “PowerShell”) and paste this, replacing YOUR_KEY with the key you copied:

curl https://api.openai.com/v1/decisions \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6-luna",
    "type": "predicate",
    "input": "Hi, my package was marked delivered 12 days ago but it never arrived. I want my money back.",
    "question": "Is this customer asking for a refund?"
  }'

Step 5 — Read the verdict. You’ll get back a short JSON answer — something like a verdict plus confidence, e.g. {"answer": true, "confidence": 0.96} (illustrative format — the exact fields follow the official docs). Translation: yes, this is a refund request, and the AI is 96% sure. That’s it. You sent ~30 words, you got one decision back, and it cost roughly $0.000004 — four millionths of a dollar.

Step 6 — Try a “choice” decision. Change "type" to "choice" and the question to “Which department should handle this: billing, shipping, or returns?” — now the AI picks one label. Change it to "score" with “How angry is this customer, 0–100?” and you get a number. Three shapes, endless sorting jobs.

Copy-paste starter: the customer-support ticket classifier

This is the exact pattern behind the money idea below. Feed each incoming message to the API with fixed instructions and get a label back:

curl https://api.openai.com/v1/decisions \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6-luna",
    "type": "choice",
    "input": "PASTE THE CUSTOMER MESSAGE HERE",
    "question": "Classify this support ticket. Choose exactly one: URGENT_REFUND (customer demands money back), ANGRY (customer is upset but no refund demand), ORDER_STATUS (asking where their order is), PRODUCT_QUESTION (asking about a product), SPAM (irrelevant or promotional).",
    "instructions": "Decide based only on the message text. If the message matches none of the categories, choose SPAM."
  }'

Your software then routes: URGENT_REFUND → owner’s phone, ORDER_STATUS → auto-reply with tracking info, SPAM → trash. You can send up to 400 messages in a single request (batch them into an array of questions per the docs), so one call can clear a whole morning’s inbox.

Make money with it: the $199/month ticket-routing plugin

The customer: small Shopify/WooCommerce store owners doing $20k–$100k/month who personally answer every customer message and lose evenings to their inbox. There are millions of them; they hang out in Facebook groups like “Shopify Entrepreneurs” and hate anything that smells like enterprise software.

What you sell: a done-for-you “inbox autopilot” — you connect their support email/chat to the Decisions API classifier above. Refund demands go to their phone instantly, order-status questions get automatic tracking replies, angry customers get flagged before they leave 1-star reviews, spam gets filtered. They change nothing about how they work; their inbox just gets quieter.

What you charge: $199/month flat, positioned against the alternative (a part-time VA at $800+/month). Your actual API cost per client: roughly $0.05–$0.30/month (ten thousand messages ≈ a few cents). That’s a ~99% gross margin before your time. What this means for you: the price is pure positioning — you’re selling reclaimed evenings, not tokens.

7-day validation plan:

1. Day 1–2: Join 2–3 Shopify/Facebook seller groups and post: “I’m testing a free tool that auto-sorts customer messages — who wants their inbox triaged free for a week?” (No website needed; a Google Form collects signups.)
2. Day 3: Take your first volunteer’s last 200 messages, run them through the classifier, and hand them a before/after report: “37 refund demands you almost missed, 112 order-status questions that could’ve auto-replied.”
3. Day 4–5: Wire it live for that volunteer (Zapier/Make + a tiny script, or even manual forwarding at first). Shadow-run it: you classify, they verify.
4. Day 6: Ask for a number: “How many hours did this save you?” Get it in writing.
5. Day 7: Offer the volunteer $199/month to keep it, and post their quote (with permission) as your first testimonial in the same groups. One paying client = validated. Three = a business.

Pitfalls and limits (read before you build)

• One model only, for now. The API currently runs exclusively on gpt-6-luna. What this means for you: if your decisions need a different model’s strengths (or pricing), you can’t switch yet — check the docs for updates before committing a big client.
• Decisions, not documents. Do not use it to write product descriptions, draft replies, or summarize articles — it returns verdicts, not prose. Pair it with a normal chat API when you need words: Decisions API sorts, the chat API writes.
• Confidence isn’t truth. A 0.96 confidence on a wrong answer still routes the ticket wrong. What this means for you: for money-moving categories (refunds, legal, medical), always keep a human in the loop for the first weeks and measure accuracy against real outcomes.
• Vendor numbers are vendor numbers. The 10× speed claim and all pricing are OpenAI’s statements, not independently verified. Budget a 2× safety margin on cost estimates you give clients.
• Data goes to OpenAI. Unless you enable Zero Data Retention (compliance tier), your inputs are processed on OpenAI’s servers. What this means for you: don’t route anything you’d be embarrassed to see in a breach — and tell clients exactly where their data goes.