Claude Sonnet costs 293x more than this

by Luis Rodrigues

What I’m building with AI + one hot take + five links worth your time.

The Take

Diogo Almeida, co-inventor of RLHF, the technique behind ChatGPT, announced a new model last week. It's called Jev, and it's a "decision model". You can't chat with it, but you can just ask it to make decisions. I got access to Jev this Friday, and I was genuinely interested in how it worked after seeing a video of it playing Doom.

You have three ways of using it: Use a Noul when the answer is yes or no. Choice when the answer is one of a fixed set of options, and Score when the answer is a position on a spectrum you can describe in steps.

I tested a bunch of decisions I normally use other models for with thoughtled.ai. It's fast, and quality matches everything I've read. The performance drops when context grows with content unrelated to the decision.

Most of the content online says something like: this is the future because most of the agents are routing, classifying, and verifying data, so you're paying frontier model prices for a classification problem.

And this time the hype is partially right. In a June paper, the authors analyzed 6,003 public n8n workflow templates and found that in 27% of the places an LLM appears, it sits between two code logic nodes.

In TypeSafe's own workflow evaluation (graded against frontier models, not ground truth), the Jev model scored 67.8%, the same as Claude Sonnet 5, at 293 times less cost and 195 times lower latency.

Constrained classifiers aren't something new, and a few open-source clones appeared this week on GitHub. Jev's real innovation isn't the architecture, but the fact they shipped a clear API that works out of the box.

Why this matters if you're building, or getting ready to: If you have a lot of decision or evaluation workflows, you should definitely try it because it will save you a lot of money.

If you're still playing with agents, go try it because this is a totally new product to help you make decisions.