“Decision model” now means two things

In September 2026 “decision model” became the name of a new kind of AI product: a small, fast model that returns a yes, a label, or a rating, and no text. Within six weeks OpenAI, Liquid, and Baseten’s lab had each shipped one.

avva uses the same two words for something else: one professional’s decisions, published for other people’s AI to read. The two share a name and almost nothing else. This article describes both, so you can tell which one a page is talking about.

The new kind: a model that only decides

• Jev, September 15. TypeSafe introduced it as “unstructured state in, typed probabilistic decisions out”. It answers with a choice, a rating, or a yes-or-no probability, and it cannot write a sentence.
• OpenAI’s Decisions API, October. First shown at DevDay at the end of September. OpenAI’s changelog dates the public beta to October 6 and general availability to October 22.
• d1-ultra, October 15. Liquid’s definition is the shortest: “Decision models answer questions about a situation with a probability for each possible answer.”
• elephant-1, October. Base Labs published open weights for “a 9B parameter decision model” under the MIT license.

What OpenAI shipped

OpenAI’s guide opens with the whole product in one sentence: “The Decisions API evaluates text, images, and JSON and returns typed answers about 10x faster than the Responses API.” You send an input and between 1 and 400 questions. Each question gets one of three kinds of answer:

• Predicate. The probability, from 0 to 1, that a condition is true. Is there visible damage in this photo?
• Choice. One of the values you supplied. Which department gets this complaint?
• Score. A rating against levels you defined, lowest to highest. How severe is this issue?

No field in the response carries an explanation. Uncached input costs $0.10 per million tokens, and the guide states: “Output tokens are not charged.” The uses OpenAI names are classifying content, routing requests, and prioritizing work.

Who actually calls it a decision model

• Liquid and Base Labs use the term for their own products.
• OpenAI does not. Its guide says “Decisions API” throughout, and the phrase “decision model” is absent from it.
• TypeSafe, which started the category, argues against the name. Its post Don’t Call It a Decision Model ends: “Jev makes judgments, code makes decisions.”

The press and the public benchmarks use the term anyway, so it has stuck.

Where the criteria come from

In every one of these products the model is general and the criteria arrive with the request. The developer writes each question, each list of values, and each level. OpenAI’s advice is to “write questions around observable criteria.”

None of them learns how a particular person decides. That is not what they are for. They answer a narrow question the same way thousands of times a minute.

The other kind: one person’s decisions, published

A Decision Model on avva is not a neural network, and avva runs no language model. It is a published record: decisions one professional made in one field, the reason for each where they recorded one, what they weigh first, where they stop, and what changes their usual answer.

Someone else’s assistant reads it over MCP in the middle of a conversation. The assistant does the reasoning and answers as itself. This guide shows how one is built.

Side by side

• What it is. The new kind: a trained model behind an API. avva’s: a published record that an assistant reads.
• A typical question. “Is this ticket urgent?” against “How would this pricing consultant handle a request for a discount?”
• What comes back. A probability or a label. Against that: past decisions and reasons for the assistant to apply.
• Who wrote the criteria. The developer, in the request. Against that: they were read off decisions a named person approved.
• Who is waiting. Software, for a fraction of a second. Against that: a person in a conversation.
• The explanation. Left out by design in OpenAI’s API. In a record of decisions, the reasons are the content.

Limits reported so far

The fast models are weeks old, and most accuracy figures come from the vendors themselves. One caution from an analyst is worth keeping. InfoWorld quotes Ashish Chaturvedi of HFS Research: “Each model’s confidence score carries its own meaning.” A threshold tuned on one model does not carry over to another.

The record kind has its own limit. It holds what one person decided, and an assistant can still apply it badly. Four checks before you rely on one are in an earlier article.

Which one you need

• Your software must sort, route, or flag things at speed: the Decisions API or one of its rivals.
• Your assistant should handle a judgment call the way a specific person would: a published record of that person’s decisions.

The two could meet. A fast model needs criteria, and someone has to write them. A threshold taken from an expert’s record is one honest source. That is our reading of where this goes. No avva feature connects the two today.

To see the record kind, open a Decision Model and put a question to it.