OpenAI's Decisions API is now open to everyone, and it's built for a job most AI tools are surprisingly bad at: answering hundreds of small, repetitive questions quickly and cheaply — including questions about photos.
The idea is simple. Instead of asking one model to write you a long answer, you send a whole batch of tiny questions and get back tidy, machine-readable results: a probability between 0 and 1, a score, or a pick from a short list of options you define yourself. No essay, no fluff, no "as an AI language model."
What happened
OpenAI took the Decisions API out of beta and made it generally available. It runs on GPT-6 Luna, accepts images as well as text, and handles a whole batch of questions in a single request — up to 400 of them, to be exact. OpenAI is pitching it as a direct answer to Jev, a rival product that also lets you fire off hundreds of questions at once and get structured answers back fast.
The plumbing is what makes it interesting. An API is just the pipe that lets your own software talk to OpenAI's models without a human in the middle. Here, you point that pipe at a batch of questions instead of a chat window.
What's actually new in the public version
• 400 questions per request, up from 200 in the beta.
• Image support on individual questions — you can pass a photo as a link (URL) or as base64, which is just a way of turning a picture into a long string of text so it can travel inside a request.
• Structured JSON input. JSON is a plain-text format for organised data — picture a spreadsheet row written out in words and brackets.
• Shared context caching, which stops the model from re-reading the same background instructions every time. It's the difference between a warehouse that re-reads your address label for every parcel and one that remembers it.
The announcement and the technical spec are both public: OpenAI's announcement and the Decisions API developer guide.
Why a "decision" API is a different animal
Most AI models behave like a brilliant freelance writer. Hand them a task and they'll produce something thoughtful, elegant, and occasionally far too creative for your liking.
A decision API behaves more like a clerk with a clipboard. You hand over a stack of forms; the clerk ticks a box on each one and hands them back. Ask it "is there damage in this photo?" and you get a number between 0 and 1, not a meditation on the philosophy of shipping.
That sounds less impressive. It's actually the thing most businesses need.
The two numbers that matter: speed and price
OpenAI says the API is up to 10 times faster than calling GPT-6 Luna through the standard Responses API, and that pricing starts at $0.10 per million input tokens. Tokens are just the chunks of text a model bills you for. An independent write-up from Analytics Vidhya and the GPT-6 Luna model page both confirm the basics of the release, and Sanity's glossary entry has a plain-English breakdown if you want the longer version.
The headline for anyone running a small operation isn't brilliance. It's speed and price.
What it means for you
Here's where it gets practical. Because you can mix text and images and ask hundreds of questions in one go, a lot of small, annoying jobs suddenly become automatable. Not "replace your team" automatable — "stop doing this by hand on a Sunday night" automatable.
At home: is that a scratch or just a smudge?
Photograph the item, ask "are there signs of damage?", and get back a probability. Useful when you're buying second-hand furniture from a stranger, checking a parcel before you sign for it, or arguing about a return. You're not asking for a verdict — you're asking for a second pair of eyes that never gets tired.
At work: routing paperwork without reading it
A photo of a document goes in, and out comes a classification: invoice, receipt, signed contract, delivery note. Then your software sends it down the right path. Offices burn a shocking amount of time on this, and it's exactly the kind of task where a fast, cheap, right-most-of-the-time answer beats a slow, expensive, perfect one.
In business: quality checks at volume
If you sell physical products, returns are a nightmare. Every returned item needs someone to look at a photo and decide: resell, repair, or write off. With a full batch of questions per request, you can push a whole day's returns through in one go and pull a human in only where the probability is genuinely murky.
For study: is the worksheet actually finished?
Upload a scan and ask a direct question — is the signature box ticked, is question 4 blank, is the date filled in. It's fast, it's cheap, and it won't get bored on page 200. Teachers, tutors and anyone marking evening paperwork will feel this one immediately.
For creators: taming the screenshot swamp
Ask whether a screenshot shows an error message, and sort your entire camera roll in one pass. Anyone who's scrolled through a phone stuffed with screenshots hunting for "that one bug report" knows exactly why that matters.
For income: a micro-service you can run for pennies
Damage checks, document sorting, screenshot triage — small sellers and solo operators pay real money for this, and the underlying cost per question is tiny. If you've been looking for a low-overhead side service, this is a cheap one to test. And if you want to see what AI can already do for free before you start paying per token, MyKreaTool has a set of free AI tools worth half an hour of your time.
How to try it right now
You don't need to be an engineer to follow the shape of this, but you will need an OpenAI developer account and a few minutes.
1. Write your question as a checkbox, not a request. "Is there damage in this image?" is good. "Tell me about this image" is not — it invites a paragraph, and paragraphs cost money.
2. Do a free sanity check first. Before you spend anything, ask the same question once in a free chat assistant. If the answer is useless there, it'll be useless through the API too.
3. Get an API key. Sign in to OpenAI's developer platform and create one. Treat it like a password, because that's what it is.
4. Copy the sample request from the Decisions API developer guide. It already shows the shape: a batch of questions, each with its own text or image.
5. Start small. Two images, five questions. Check that the numbers look sane. Then scale up toward the full batch limit.
6. Turn on context caching so your background instructions aren't re-charged on every single call.
7. Log everything. Save the probabilities, spot-check 20 of them by hand, and only then trust the batch.
One honest note on "free": there's no free tier for the API itself. But at $0.10 per million input tokens your first experiments cost pennies, and images add tokens, so keep your test photos modest. The full details live on OpenAI's Decisions API page.
Upsides and what changes
• Boring tasks get cheap. Decisions that were never worth automating suddenly are, because the unit cost is tiny.
• Scale stops being a project. A big batch per request means fewer round trips, less glue code, and less waiting around.
• Images stop being a special case. You can mix "is this damaged?" and "which category is this?" in the same batch.
• Caching changes the maths. Reused context is cheaper, which matters when you're running the same instruction thousands of times a day.
• The exit door stays open. Because answers come back as numbers and categories, you can slot a human review step in wherever confidence is low.
The bigger shift is psychological. For years, companies treated AI as a writer. This nudges it toward being a classifier that happens to read photos — and classifiers are the workhorses of real operations.
Limitations
Be honest with yourself about what this is. A decision API is a fast, cheap sorter, not an expert — a probability of 0.8 isn't a guarantee, and it's still your job to decide which number you'll act on. Blurry photos, bad lighting and unusual documents will all drag accuracy down, and the speed and price figures are OpenAI's own claims rather than an independent benchmark. There's no free tier, you'll need some developer help to wire it in, and you're still uploading photos of real documents and real customers to a third party — a conversation your lawyer should be in on. Finally, there's a ceiling on how many questions fit in one request; bigger jobs mean batching, and batching means code.
Conclusion
OpenAI's Decisions API isn't the flashiest thing the company has shipped this year. It's the most immediately useful if you've got repetitive eyeball work: checking photos for damage, sorting paperwork, triaging screenshots, at a price per question that rounds to nothing.
Here's your one action for today: pick one thing you or your team looks at 50 times a week, and write it down as a yes/no question. That single sentence is the whole starting point — and if you genuinely can't phrase it as a yes/no, that's useful information too, because it means the job isn't as simple as it feels. Once you've got the sentence, you'll know whether this API deserves an hour of your week.
Comments 0