What is a decision model?
Some AI models never write a word. They tick a box, tell you the odds, and sort a million messages for about $3.52.
On this page
“Hi, my order #4471 arrived this morning and the tablet screen is cracked right across the corner. I have already written twice this week with no reply. I fly out on Friday and need this sorted before then. Please send a replacement or refund me.”
A customer writes to say their tablet arrived cracked and they fly out on Friday. Before anyone answers, three small questions have to be settled. Is it urgent? Which team gets it? How upset is this person?
Most software now hands those questions to a chatbot-style model, which writes a little paragraph that a program then has to pick apart. A decision model skips the writing. You give it the allowed answers, and it tells you how likely each one is. That makes it the cheaper and faster tool for sorting, routing and checking, and the wrong tool for anything that needs words. This post explains how it works, shows a real run with the bill, and ends with a simple test for when to use one.
A decision model picks from your answers and never writes
A general AI model, the kind behind ChatGPT, Claude or Gemini, produces text. Ask it anything and it writes back, one piece of a word at a time. A decision model cannot do that. It has no way to produce a sentence.
What it takes is three things: something to look at (a message, a document, a photo), a question, and the list of answers you will accept. What it returns is a probability for each answer. The vendors have settled on three kinds of question, and OpenAI’s guide lays them out plainly: a yes-or-no check, a choice among options you list, and a score on a scale you define.
The idea is older than the name. Researchers were already ranking documents by the probability of “true” and “false” in 2020, as the startup Sqwish notes in its own history of the field. The category took off in September 2026, when TypeSafe AI released a model called Jev and described it as a “System One” model, after the fast, intuitive kind of thinking in Daniel Kahneman’s Thinking, Fast and Slow. Cloudflare, OpenAI, Microsoft, Liquid AI, Perplexity and Google all shipped their own within two months. OpenRouter’s list held 43 entries of this type when we read it on November 17.
Four models gave one message the same routing
We sent the cracked-tablet message from the top of this post to four decision models, with the same three questions each time. The questions were: does this need a reply today, which of four teams should handle it, and is the customer calm, annoyed or angry. The full message is in the panel at the top of the page.
For technical readers, this is the whole request. The same body works on every decision model OpenRouter lists, and the reply is JSON with one entry per question.
POST https://openrouter.ai/api/v1/systemone
{
"model": "liquid/d1.2",
"state": "Hi, my order #4471 arrived this morning and ...",
"questions": {
"urgent": {
"type": "noul",
"instructions": "Does this message need a reply today?"
},
"team": {
"type": "choice",
"instructions": "Which team should handle this message?",
"criteria": {
"billing": "Charges, invoices and payment problems",
"returns": "Damaged, wrong or unwanted items",
"shipping": "Delivery timing and tracking",
"technical": "Help using a working product"
}
},
"mood": {
"type": "score",
"instructions": "How upset is the customer?",
"criteria": ["Calm", "Annoyed", "Angry"]
}
}
}All four agreed on the two questions that drive an action. Each put the chance that the message needs a reply today at 90% or higher, and each sent it to Returns with at least 98% probability. A program can act on that without a person reading anything.
The mood question is more interesting. Three models leaned toward “annoyed” and one split it almost evenly with “angry”. None of them was sure, and all of them said so in numbers. When we asked a chat model the same thing, it answered “Annoyed” and gave no sign that it was a close call.
It reads once and scores every answer at the same time
The speed comes from what the model skips. A chat model writes its answer in small pieces called tokens, and each token needs its own trip through the model. Many also think before they write, which costs more tokens. In our run the chat model spent 69 tokens thinking and 24 writing to produce one line of JSON.
- 1Reads the message and your instructions
- 2Thinks it over (69 hidden tokens in our run)
- 3Writes the reply one token at a time (24 more)
- 4Your code reads the text and hopes it fits
- 1Reads the message and your questions
- 2Scores every allowed answer in the same pass
- 3Returns numbers your code can use as they are
A decision model reads the input once and stops. Cloudflare’s launch post for its Clef models describes the mechanism: the model makes one pass over the input, then scores the valid choices in parallel, with “no intermediate text to generate token by token”. Asking more questions is nearly free for the same reason. Liquid AI measured its small open model at 8 milliseconds a question on an NVIDIA RTX 4090 graphics card, and found that three questions took only 1.3 times as long as one.
Those are numbers for the model alone, on local hardware. Over the internet the network takes most of the time. Our four decision calls took 1.0 to 1.4 seconds each from our desk, and the chat model took 3.9 seconds. Microsoft’s announcement of its own decision model explains why this matters in automation: add 100 milliseconds to each of 20 decisions in a row and the whole job is two seconds slower.
The number beside the answer is the useful part
A plain “yes” forces you to trust it or not. A “yes, 97%” lets you set a rule. You can act on anything above a line, ignore anything below another, and send the uncertain middle to a person.
This only works if the numbers are honest. When a model says 80%, it should be right about 80% of the time. That property is called calibration, and it is what these models are trained for. Chat models are poor at it: the same prompt can give a different answer each time, and a confidence figure a chat model writes out is itself just more generated text.
Calibration can be measured, and it varies. The independent Decision Index from Artificial Analysis builds it into the score, so a confident wrong answer costs more than a hesitant one. Liquid’s own tests put the gap between stated and actual accuracy at 0.8 points for its best model and 7.2 points for OpenAI’s. That is a vendor’s measurement, so treat it as a claim. The practical advice in OpenAI’s guide holds for every model: set your thresholds from examples you have already labelled, and weigh what a wrong yes costs against a wrong no.
The bill counts what the model reads, not what it says
Chat models charge for input and for output, and output costs more. Decision models have no output to charge for. Every one on OpenRouter’s list bills input tokens only.
On our small message the totals were tiny either way, so the ratio is the thing to read. Liquid’s d1.2 handled the three questions for $0.0000035. The chat model cost $0.0000583, about 17 times more. At a million messages that is $3.52 against $58.30.
The gap is not always that wide. OpenAI’s decision endpoint cost $0.0000455 for the same request, only a little under its chat model, because it counted more input tokens. TypeSafe’s launch post claims savings of more than 400 times in its own tests, and says in the same post that it expects those figures to be “on the higher end of real world gains”. The saving grows with how much the chat model would have written or thought. Our post on cost per task explains why the per-token price alone can mislead.
It cannot explain itself, and it can still be wrong
A decision model never tells you why. If you need a reason, a summary, a draft reply or a value that is not on a list you wrote, you need a model that writes. OpenAI’s guide says the same and points to its text API for anything that needs “a written explanation”. A few vendors are widening the format (Liquid’s newest models can read a figure off an invoice or mark passages in a text), but the answer is still a typed value and never prose.
It also cannot work through a problem in steps. On the Decision Index, read November 17, the best dedicated decision model scores 74.8. Two small general models given their maximum thinking effort score 76.5 and 75.7, and the same two at low effort drop to 67.2 and 68.8. Thinking still helps on hard questions. It just costs time and money that a simple sorting job does not need. Our guide to reasoning models covers that trade.
One claim needs care. TypeSafe says its model “can’t hallucinate”. That is true in a narrow sense: the model cannot invent an answer outside your list. It can still pick the wrong item on the list, and a 98% can be wrong 2 times in 100. A badly worded question or two overlapping options will produce confident nonsense from any of them.
Use one wherever the answer is a box to tick
The test is short. If you could print the possible answers on a form before the question is asked, a decision model fits. If the answer is something you would have to read, it does not.
That covers a lot of everyday work: sorting an inbox, flagging reviews that need a reply, checking whether a photo shows damage, deciding which of ten documents answers a question, grading a draft against a rubric. Sqwish walks through a support assistant in which seven of eight steps choose among options the software already holds, and only the last one, writing the reply, needs a model that writes.
The quality of the answer depends on the question, much as it does with a prompt for a chat model. OpenAI’s guide gives three rules that apply to every model here. Ask about things that can be observed in the input. Put separate concerns in separate questions, so “is it urgent” and “is it angry” never share one answer. Give each option a meaning that does not overlap with its neighbours. When one decision depends on another, send two requests and let your code use the first answer to choose the second question.
Then test it before you trust it. Take 50 items you have already sorted by hand, run them through, and compare. The misses will show you whether to reword a question, add an option or move a threshold. This takes an afternoon and costs a fraction of a cent.
The pattern that follows is to use both. Let a decision model do the sorting, and call a chat model only for the items that need words. This is how an AI workflow stays cheap as it grows. In CSuite, the Decide step in a workflow does this, with a hosted model or with one that runs on your own computer. Several of these models have open weights, Cloudflare’s Clef and Liquid’s d1 among them, so the sorting can happen without the message leaving your machine.
So the next time an AI tool takes four seconds and a paragraph to tell you “yes”, remember the cracked tablet. Three answers, with honest odds on each, cost $3.52 per million messages. Save the writing models for the writing.


