OpenAI now sells a yes-or-no as its own API call. Split each process into decisions and writing

Look at almost any process a small company runs and most of the steps are not writing. They are choices: which queue this ticket goes to, whether this invoice matches the order, whether this lead is worth a call. Our guide to choosing what to hand to AI first says to start with frequent, rule-bound, easy-to-check work. On October 6, OpenAI put a price on a decision call for that kind of step.

What shipped

OpenAI’s developer community announced that the Decisions API is in public beta for all developers. According to the announcement, it runs on GPT-6 Luna, takes text and images, and returns one of three kinds of answer:

• Predicates: an estimate of the probability that a statement is true.
• Choices: one option from a list you define, with confidence scores.
• Scores: an evaluation of an input against a numeric range.

OpenAI says it makes these decisions up to 10 times faster than the same model through the Responses API. The price is $0.10 per million input tokens, and you pay only for input: there are no cache-read, cache-write or output-token charges. The speed claim and the price are company-reported, and this is a beta.

Why this matters to a lean company

On a general chat model, a routing step and a client email are the same kind of call. You ask the model to write, then pull a label out of its answer, and you pay for words nobody reads.

A separate decision call changes how a process should be drawn. The steps that end in a label, a yes or no, or a number can go on a cheap, fast call. The steps a client or employee actually reads keep the larger model. The confidence score adds one more thing: a place to draw the line between what the machine settles and what a person checks.

The decision rule

Go through one process step by step and ask one question of each: what does this step hand to the next one?

1. A label from a fixed list (queue, category, priority). That is a Choice.
2. A yes or no (does this match, is this complete, is this allowed). That is a Predicate.
3. A number on a scale (lead fit, urgency, risk from 1 to 5). That is a Score.
4. Words someone will read. That is writing. Leave it on the model you use now.

If a step hands over both a decision and words, split it in two. Decide first, then write only for the cases that need it.

A worked example

Take an illustrative support inbox with 3,000 messages a month. Each message, with the instructions and the list of queues, comes to about 800 input tokens.

• Routing all 3,000 is 2.4 million input tokens.
• At the company-reported $0.10 per million, that is about $0.24 a month.
• With no output charge, the label itself adds nothing to that figure.

These volumes are made up for the arithmetic; your own token counts will differ. The point is the order of magnitude. At this price, the cost of the call stops being the question. The questions that remain are whether the answers are right and what happens when the model is unsure.

A test you can run this week

You do not need to rebuild anything to find out if this works for you. You need 50 past cases where you already know the right answer.

1. Pick one decision step from the rule above. Routing is the easiest to start with.
2. Pull 50 past cases and write down the answer a person gave each one.
3. Run them as a Choice with your real list of options, and record the choice and the confidence score.
4. Sort by confidence, highest first. Find the point above which every answer matched the person’s.
5. Set that as your line. Above it, the call decides. Below it, a person decides, as they do today.
6. Count the share above the line. That share is the work you could hand off, not the whole inbox.

Fifty cases is a small sample, so treat the line as a starting guess and recheck it on new cases each month. Keep the old path running while the API is in beta, and do not put a step that touches money, contracts or client data on it until you have checked where that data goes.

What to do this week

• List the steps in one process and mark each as Choice, Predicate, Score or writing.
• Split any step that both decides and writes.
• Run the 50-case test on one decision step.
• Write down your confidence line and who handles the cases below it.
• Keep your current setup as the fallback until the beta ends.

This is the kind of sorting we do with owners in a Build month, where each month the client picks up to five processes to build with our coaches. Prices and betas will change. Knowing which of your steps are decisions does not.