System 1 Models: The Opposite of Reasoning Models

Watch on YouTube

In my bookmark app, Markwise, every saved link gets one category. I used an LLM for that. It took about 2 seconds to pick one word, because it wrote around 100 tokens of thinking first. In my test, 16 of the 2,000 answers came back empty: the thinking used up the token cap before the model said the word.

Then I read about a new kind of model: System 1 models.

System 1 vs System 2

The name comes from Kahneman's Thinking, Fast and Slow. "2 + 2" you just know. That is System 1. "17 × 24" you sit down and work out. That is System 2.

A bookmark is the same. You see a link to a pasta recipe and you know it is "cooking". Reasoning models are built for System 2. These new models only do System 1: you give them the situation and the options, they give back a probability per option, in one step. No text, no thinking tokens.

Two of them are out now:

Cloudflare calls Clef a "decision model". Their post says Clef-flash is 13x faster than Jev (39 ms vs 524 ms, their medians). On their own test against an LLM, the gap is about 2x.

The test

I took 2,000 bookmarks from my own app. The old LLM had saved 5,657 different category names for them, so step one was a cleanup. I let the models do it themselves: Clef merged 5,657 names into 64 in 5 rounds, in 28 seconds, for $0.13.

Then every model picked one of the 64 categories for each bookmark, from my laptop. The LLM baseline was GLM-5.3-Flash.

Speed and cost

From my laptop, Jev was the fastest, about 2x faster than Clef. That is the opposite of Cloudflare's claim. My guess is their setup or region; I can't see it from here. Clef cost about 7x more than Jev.

The GLM cost is at async rates. The realtime price is not published, so treat it as an estimate.

Are they right?

There is no clean answer key. The saved labels came from the old LLM, and they are messy. So these numbers measure agreement, not truth:

All four land around 40%, so no winner on the top pick. Jev and Clef picked the same category 75% of the time.

Can I trust the probability?

This is the part I care about most. A System 1 model gives a probability, so my code can decide what to do when the model is unsure.

Jev says it is sure too often. Clef rarely says it, but when it does, it is right more often.

What I take from it

For the many small choices in an app, a System 1 model goes first. My setup: Jev for every bookmark, and when the answer looks unsure, a second opinion from Clef or an LLM. Save the reasoning model for the jobs that need thinking.