OpenAI Decision API: What It Costs and How to Sort Photos
OpenAI's decision API picks one label per item, 128 per request, at $0.10 per million input tokens with free output. See how it sorted 5,600 photos.
For Engineering, Operations and Finance

What is OpenAI's decision API?
OpenAI's decision API is a way to have an AI model look at an item and pick one label from a list you provide. You can send up to 128 items in a single request. OpenAI says these decisions run up to 10 times faster than the regular API with GPT6 Luna. Pricing is $0.10 per million tokens going in, and the answers coming out cost nothing. In one demo, a creator used it to sort 5,600 iPhone photos into screenshots, blurry photos, and keepers.
That is the whole idea: not a chatbot, but a fast, cheap sorter.
TL;DR
- The decision API picks one label from your list for each item, instead of writing free-form text.
- A single request can carry up to 128 items, so a big pile of photos becomes a small number of calls.
- OpenAI claims decisions run up to 10 times faster than the regular API with GPT6 Luna. That is the company's claim, not an independent test.
- Pricing is $0.10 per million input tokens, and output is free, so the cost is driven by what you send in, not what comes back.
- A demo sorted 5,600 photos into screenshots, blurry shots, and keepers, then filed them into albums.
- Developers can try it in OpenAI's decisions playground before building anything.
- Sending personal photos to any outside service is a privacy choice. Think about it before you run your whole library through.
How does the decision API work?
A normal chat request to an AI model ends with the model writing text. You ask a question and get paragraphs back. That is flexible, but it is slow and wasteful when all you need is a category.
The decision API narrows the job. You give it two things: the items to judge, and the list of allowed labels. For each item, it returns one label from your list. It is not inventing an answer. It is choosing among the options you set.
This matches a task people have long done by hand or with older machine learning tools: classification. Is this email spam or not? Is this expense travel or meals? Is this support ticket urgent? Is this photo a screenshot or a real picture? The new part is that a general-purpose model can handle the call without you training a custom classifier first.
Batching is the other half. Instead of one request per photo, you send up to 128 at once. For a library of 5,600 photos, that is a few dozen requests rather than thousands. Fewer round trips means less waiting and simpler code.
How much does the OpenAI decision API cost?
Based on what has been shown publicly, the pricing has two parts:
| Part | Price |
|---|---|
| Input (what you send) | $0.10 per million tokens |
| Output (the labels you get back) | Free, no output token cost |
A token is the small chunk of data a model reads. For text, it is roughly a short word or word fragment. Images are also converted into tokens, though the exact count per photo depends on how the service processes them, and the source material does not say.
So here is how to think about it, using a made-up number. Suppose each photo counted as 1,000 tokens. That is an assumption for illustration, not a published figure. Then 5,600 photos would be 5.6 million tokens, and at $0.10 per million, the input would cost about 56 cents. If each photo counted as 10 times that, it would be about $5.60. In both cases, the cost of sorting a whole phone library would be small.
The practical point stands even without the exact token count. Because output is free, you can ask for decisions on a lot of items without worrying that the answers add to the bill. Check OpenAI's current pricing page before you plan anything at scale, since prices and model names change.
Is the "10 times faster" claim real?
OpenAI says decisions run up to 10 times faster than the regular API with GPT6 Luna. Two things are worth keeping in mind.
First, "up to" means best case. Your speed will depend on what you send and how big your batches are.
Second, this is OpenAI's own claim, repeated in a demo. I have not seen an independent benchmark. If speed matters for your project, run a small test with your own data and time it.
The design does give a plausible reason for a speed gain. Returning a single label from a fixed list is a much smaller job than writing a paragraph, and batching 128 items removes a lot of overhead.
How can you sort photos with the decision API?
The demo is a good template. The creator built an app that looked at each photo on an iPhone, asked the decision API to pick one label from a list, and then filed the photo into the matching album. The steps look like this:
- Define your labels. In the demo, the buckets were screenshots, blurry photos, and photos worth keeping. A "too far away" label also showed up, used for photos of a tiny person up on stage at a tech conference.
- Group photos into batches of up to 128. Each request carries a batch plus your label list.
- Send the batches. The API returns one label per photo.
- File the results. Your app moves each photo into the album that matches its label.
- Review the edge cases. Skim the albums, especially "delete" candidates, before removing anything.
The demo caught accidental pocket shots, too. That is a good example of a label you would never think to search for by hand but that a model can spot instantly.
A label prompt you can adapt
If you try this in the decisions playground, start with plain, mutually exclusive labels and a clear rule for each. Here is a template to adapt:
Task: For each photo, choose exactly one label.
Labels:
- screenshot: an image of a phone or computer screen
- blurry: out of focus or motion-blurred, not usable
- accidental: pocket shot, floor, finger over lens, or black frame
- too_far: the main subject is too small to see clearly
- keep: sharp, well-framed, worth keeping
Rule: if more than one label fits, pick the first one in the list above.
The tie-break rule matters. When labels overlap, the model has to guess which one you meant, and your results get inconsistent.
Can it do more than pick a bucket?
The creator also ran a second pass. The phone showed 58 different trips since 2018, and the app had the AI score every trip photo on how sharp and well framed it was. That produced one album with the best shots from each trip.
The source does not explain exactly how the scoring was set up, so I will not guess at the technical details. The takeaway is that the same pattern, many items in and one judgment per item out, covers ranking as well as sorting. You can sort first, then rank within each group.
What else could you classify with it?
Photos are the fun demo, but the pattern applies anywhere you have a big pile of similar things and a short list of categories. A few examples from ordinary jobs:
- Support inboxes: label each message as billing, bug, feature request, or cancellation.
- Expense reports: label each receipt by category.
- Resumes or applications: a first pass of "meets minimum requirements" or "does not," with a person making every final call.
- Survey comments: positive, negative, or neutral, plus a topic label.
- Product catalogs: match each listing to the right category.
For anything that affects people, such as hiring or credit, keep a human in the loop and check for bias in how the labels are applied.
What should you watch out for?
- Privacy. Your photos would be sent to an outside service. Read OpenAI's data terms and think about which photos you are comfortable uploading. Family and children's photos, IDs, and anything work-related deserve extra thought.
- Mistakes. A model will misjudge some photos. A slightly soft image might be marked blurry, and an artistic shot might be flagged as a bad frame. Do not auto-delete. Move "bad" photos to a review album first.
- Label design. Vague or overlapping labels give vague results. Keep each label distinct and test on 50 or 100 items before running thousands.
- Details that may change. Model names, limits and prices on a new API can shift. Confirm the current numbers in OpenAI's documentation.
Could you build your own sorting tool?
The decision API is a building block, not a finished app. You still need somewhere to hold the items, a screen to review the results, and a way to correct the labels. If you would rather not wire that up by hand, Remy can build it from a plain-English description. For example, you could describe a triage tool for a shared inbox, a receipt sorter, or a review queue with a database, sign-in, and a page where your team accepts or fixes each label. Remy apps can use nearly 400 AI models, so the classification step can sit inside the app. You own the resulting code and keep a review screen of your own.
Frequently Asked Questions
What does the OpenAI decision API do?
It looks at each item you send and picks one label from a list you provide. You can include up to 128 items in a single request. It suits sorting and classification jobs, such as separating screenshots, blurry photos and keepers in a photo library.
How much does the OpenAI decision API cost?
According to the demo of the new API, input costs $0.10 per million tokens, and output is free. The total depends on how many tokens your items use. Check OpenAI's pricing page for current numbers before you commit to a large job.
Is the decision API really 10 times faster?
OpenAI says decisions run up to 10 times faster than the regular API with GPT6 Luna. That is the company's claim, and "up to" describes a best case. Test it on your own data if speed matters.
How can I try it without building an app?
Developers can use OpenAI's decisions playground to experiment with labels and sample items. It is a good place to test your label list on a small set before writing any code.
Is it safe to upload my whole photo library?
That depends on your comfort with sending personal images to a third party. Review OpenAI's data policies, start with a small batch of non-sensitive photos, and never delete anything automatically based on the labels alone.
How we source, write, and correct our articles. Editorial standards
Related articles

How Much Should You Spend on AI? Hormozi's Budget Lesson

OpenBot vs Grok Bot: Which AI Agent Should You Use?

How Do PandaCode Council Reviewers Check AI Agent Work?
