What happened
On 22 October OpenAI took the Decisions API out of beta and made it generally available. It is a separate endpoint that runs on a single model, GPT-6 Luna, and does not write text. You ask it closed-ended questions and it returns one of three result types: the probability that a statement is true, one option chosen from a list together with a confidence and a distribution, or a rating on an ordered scale with a probability for each level. The API was first shown at DevDay on 29 September, and the public beta opened on 6 October.
Pricing: USD 0.10 per million input tokens, USD 0.01 for cached input, and nothing for output. One request can carry up to 400 questions, twice the beta limit. Regional processing works in the US and in Europe (the EEA plus Switzerland), and zero data retention is available to eligible customers. OpenAI says answers arrive many times faster than when the same question goes to Luna through the Responses API. The product lead told TechCrunch about a drop from one to three seconds down to 100–180 milliseconds. Those are the vendor's numbers, not an independent measurement.
The documentation draws its own boundaries: extracting fields and building objects against your own JSON schema is still a job for Structured Outputs, and calling tools is still function calling.
Our read
We flag this as our interpretation. The headline is about price, but the input rate is exactly the same as a regular Luna call. The difference on the bill is the missing output charge, and in classification the answer tends to be short anyway. Anyone who works out the savings from the price list alone may be disappointed. When Haiku 5.5 launched we wrote that the headline rate is not the price of a task. The same holds here: OpenAI's price list has higher rates for very long context and a surcharge on regional processing for newer models. Before you cost out calls from the EU, check which of those lines apply to your traffic.
The real change is elsewhere. Many AI-assisted business processes contain a step that writes nothing and only decides: which team gets the ticket, whether the document is complete, whether the complaint is urgent. Until now that step usually meant the model generated a word or a JSON blob and code parsed it. You got one answer and no sense of how sure the model was. Now the same step returns a number. A number lets you set a threshold: above it, automation; below it, a person. That threshold can be written down, tested against history and changed when the cost of an error changes. With a text answer you cannot do that.
A second observation, also interpretation: a large share of processes marketed as agentic are in practice a few decisions like these wired together with ordinary code. An endpoint that only makes decisions, quickly and without text, strengthens the case that many of those processes do not need an agent. They need well-set thresholds.
There is a catch. The probability a model returns is not the same as the probability of being right on your data. We expect, and this is speculation, that on unusual documents, in Polish or in a narrow industry, the model's confidence will drift away from its real accuracy. Until you check it against your own history, the threshold is a guess with two decimal places. On top of that, this is one vendor's closed endpoint on one model. The response format will not carry over to another provider as is.
Why it matters
Private Equity
In portfolio companies that automate customer service or document flow, ask where the escalation decision sits in the process and who set it. If the answer is "the model judges for itself", that is not a controlled process. A threshold written into configuration and checked against data is something you can show in due diligence.
Enterprise
Classification steps that currently run through text generation are natural candidates to move, provided EU regional processing and the retention terms fit your data policy. Start with a high-volume step where the cost of an error is known. Log the numeric result together with the decision, because when an auditor asks, a trail reading "0.62, below the 0.8 threshold, passed to a person" is worth more than the label alone.
SMB / mid-market
You do not need to build anything to benefit. With ready-made AI products that sort documents or tickets, ask the vendor whether there is a confidence threshold, who sets it and what happens to the cases below it. If there is no answer, every decision is automatic, including the ones where the model was guessing.
One move for this week
Pick one classification step in your process and take 200 historical cases where you know the correct answer. Run them through the Decisions API or through the model you already use, and record the confidence for each answer. Sort the results by confidence and see below which value the errors start. That is your first threshold, and only once you have it should you work out whether switching endpoints pays off. Describe your case: mailto:[email protected]?subject=Rozmowa%20z%20Aurora%20AI