OpenAI's Decisions API picks from options you set, not free text: cheap for sorting tasks
OpenAI's Decisions API picks from options you define, now including voice and images, and charges only for input, which makes sorting, routing and approval features cheap at volume
Today in AI (2026-10-23)
- OpenAI's Decisions API is now open to all, with no charge for its answers: Founders can add fast, cheap AI choices like sorting, routing or approving to their product without paying per word of output.
- Anthropic's Claude now hides an invisible watermark in all the text it writes: Text your product or team generates with Claude can now be identified as AI-written, which matters for content, marketing and EU customers.
- Anthropic hands its Agent Skills standard to the Linux Foundation: Skills built for Claude are more likely to work across other AI providers, which lowers the risk of being locked in to one vendor.
OpenAI's Decisions API is an AI service that picks one answer from a list you give it, instead of writing a paragraph. If your product sorts, tags, routes, or approves things, it could be a much cheaper and faster way to do that work than a chat-style AI.
It became available to all customers on Thursday, with new voice and image inputs. OpenAI charges $0.10 per million units of input and nothing for the answers it sends back, according to TechCrunch's report.
What happened
OpenAI is entering a market that a startup called TypeSafe created with its Jev model. TypeSafe was recently valued at $7.6 billion. Jev showed that a model built only to choose between options could be quick and cheap, and developers started swapping it in for chat-style AI in parts of their products.
OpenAI's answer is the Decisions API. Nikunj Handa, who leads product for OpenAI's APIs, said the team took its existing ChatGPT Luna model and reconfigured it to run faster and cost less. He said response times dropped from one to three seconds to roughly 100 to 180 milliseconds.
What it is, in plain words
A chat-style AI, like ChatGPT, writes free text. You ask a question and it composes an answer, word by word. That is flexible, but slow and more expensive, because the model has to generate every word.
A decision model works differently. You hand it an input and a fixed list of options, and it picks one. It also gives a confidence rating, meaning how sure it is about its pick.
Think of a mailroom clerk. A chat AI is a clerk who reads each letter and writes you a memo about it. A decision model is a clerk who reads the letter and drops it into one of five labeled bins, then tells you how sure they were. For a lot of business work, the bin is all you needed.
Why the pricing matters
Most AI services charge for what you send in and for what the model writes back. The Decisions API charges only for input, at $0.10 per million units of input. (Those units are small chunks of text, roughly word fragments, that AI services use to measure usage.) The answers cost nothing.
For a sorting job, the answer is tiny anyway: one label from your list. So the real cost is in reading the input, and that part is cheap. The effect shows up at volume. If you classify thousands of items a day, a low price per item adds up to a very different monthly bill than a chat model would.
I won't put a monthly dollar figure on it, because that depends on how much text each item contains. That is exactly the number to ask your developer to estimate (more on that below).
Which product jobs it suits
The fit is any job where the answer is one of a known set of options:
- Tagging support tickets by topic or urgency
- Routing new leads to the right salesperson or follow-up path
- Approving or flagging user-submitted content
- Classifying photos, for example "damaged" or "undamaged," now that images are supported
- Sorting voice notes or short voice messages by intent, now that voice is supported
It is a poor fit when you need a written reply, a summary, a draft email, or a conversation. Keep a chat-style model for those. Many products will end up using both: a decision model at the front door to sort things, and a chat model behind it for the few items that need real writing.
Speed is the other half of the story. Speak, which builds AI language tutors, cut the time between actions from 1.4 seconds to 0.13 seconds using the API. If your product feels sluggish because it waits on AI, that matters as much as cost.
What is real and what is unproven
Be careful with the rankings. TechCrunch reports that it is not clear what makes a good decision model beyond "it works for you." Public benchmarks (standardized comparison tests) place OpenAI's model eighth and fourteenth on two indices, while other models rank higher, and TypeSafe's own CEO urges people to distrust benchmarks and test on their own work. OpenAI's pitch is strength with images and better calibration, meaning its confidence ratings are more trustworthy. That is a claim from OpenAI's product lead, so test it yourself.
I have not used the Decisions API, and I won't pretend otherwise. It was released this week. Treat everything here as a reason to run a test, not a verdict.
What to ask your developer or agency
If your product has a sorting, tagging, or routing step today, put these three questions to whoever builds it:
- How does it do on our real data? Take a few hundred real items you have already sorted correctly. Run them through your current setup and through the Decisions API. Compare accuracy, speed, and cost side by side. Pay attention to the confidence rating: does it flag the items it gets wrong?
- What will it cost per month? Ask for an estimate from your actual volume and typical input size. Because answers are free, the estimate only needs the input side. If the number comes back vague, that is a warning sign.
- Can we switch providers later? TypeSafe's Jev, OpenAI's version, and models from others are all competing here, and the leaderboard is changing weekly. Your product should call the AI through one swappable piece of code, so changing providers is a small job. I wrote about why this matters in why your product shouldn't rely on one AI provider.
Who can ignore this? If your product has no sorting or routing steps, and the AI mostly writes or chats, nothing here changes your week.
Bottom line
The Decisions API is a cheap, fast way to make one-of-N choices, and charging only for input makes it attractive at volume. It is also a young category with no clear winner, so test on your own data and keep your options open.
This week: list every place in your product where a person or an AI picks from a fixed set of options. Pick the highest-volume one and ask your developer for a side-by-side test on real examples.