What is a System 1 model?
We are all familiar with the chatbots and GPT models of the day, but there's more to AI than just ChatGPT and Claude. Just a few weeks ago, an OpenAI researcher named Diogo Almeida came out with a new type of model to show the world. This new type of model is what some people have called a classification model. But in the new world, the internet is calling it a System 1 model. The main difference between the traditional System 2 model and the System 1 model is the amount of time that it takes for reasoning and its output.
The way a typical AI model or GPT works is it predicts the next token, or word. This was commonly used and seen before ChatGPT with autocomplete on your phone, which helps you best predict the next word to type out in a text message to a friend. ChatGPT works in a similar manner but at a much larger scale, predicting entire sentences, paragraphs and even complete essays based off of a single prompt. System 2 models are used to providing you with a page or two of output, but that is not always needed.
What if you want to know if your email should be classified as an invoice, deal, or potential customer? You don't need a GPT model to provide you with a paragraph of its best guess; you'd much rather have a simple guess or probability. That's exactly what the new System 1 models of our day are here to do.
Diogo first announced Jev to the world about a month ago, on September 15th, 2026. Since then, more than 70 different System 1 models have been released, including models such as Microsoft Decision, Amazon Strands Decider 2B, Cloudflare Clef, and the OpenAI Decisions API.
Why should we care?
That's a great question. I think that these models are a great fit for the tasks that probably didn't need a full-powered LLM in the first place. There are plenty of tasks that don't need a full paragraph of response every time they are run. Jev is super efficient and can help companies make these decisions quicker and more efficiently.
Also, sometimes models are not efficient and burn through tokens at a ridiculous rate. “Tokenmaxxing” does not actually lead to real results. For the past year or so, companies that have been trying to track AI adoption through token spend have been doing it wrong. I feel like most of these efforts are led by CFOs who feel pressure to use AI, but don't understand that a token used does not equate to productivity, but rather to busyness. There are plenty of ways that an employee can look busy, and tokenmaxxing is one of the newest additions to the list.
System 1 models are also way cheaper. Jev costs $0.042 per million input tokens. That is anywhere from 238x to 445x cheaper to operate than frontier LLMs or general chat models, depending on the specific GPT/LLM. For that matter, companies can utilize System 1 models whenever and however it makes sense.
I don't know what the life expectancy of these classification models looks like. Will they make a single headline in 2027? Probably a few. Are they changing the game of AI? I'd say so. These models are the first of their kind, and the frontier adoption makes it seem like they are here to stay. My gut says that 70 models aren't released in a month because of a short fad; rather, there is a wave, and people don't want to miss it.
To learn more, here's Martin from IBM.
And here's Jev's CEO at AI Engineer.