Jev never writes a word. You hand it a document and a question, and it hands back an answer.
Today we're looking at a very different kind of AI lab: TypeSafe, and its decision model, Jev.
Jev never writes a word. You hand it a document and a question, and it hands back an answer. Yes or no, a probability, or a multiple choice pick, each with a confidence value.
The price is what makes it compelling. Just over 4 cents per million input tokens, and output is free. That's an incredible value for making your main AI better.
Jev makes three kinds of decisions. A yes or no, with the probability on each. Give it a document and questions about that document, and it answers them. A multiple choice: give it the information and the choices, and it comes back with one, again with a confidence score. And a rubric score: give it a document and the rubric, and it grades against it, each grade with its own confidence value.
So where does an AI that can't write create value? There are a lot of creative uses, but two stand out. The first is classification. Give it an item and a list of options, and it tells you which one the item belongs in. That's useful in a lot of workflows. The second is validation. Does this LLM output match these criteria? Does this LLM output actually exist in this document? It's an incredibly fast way to fact-check AI output.
How much traction does it have? TypeSafe launched Jev on September 15 with a $40 million seed at a $200 million valuation. On October 9 it raised an $870 million Series A at a $7.5 billion valuation. That's 24 days and a 37.5x step up.
Watch how fast the category fills up. Within weeks of Jev's launch, OpenAI shipped a Decisions API, AWS shipped a decider model, and Cloudflare launched its own take. Three big players are already building for the exact problem Jev solves.
Where does Jev fail? Anywhere it has to work out an answer instead of picking from a set of choices. Ask it to reason beyond its own context and the answer quality degrades.
I like Jev, and I use it in a few places: checking AI answers as they come out of interview transcripts, and as a classifier. I've watched it improve my workflows. In my classification workflow, it took overall efficiency from 87% to over 95%. A big jump from a small change. Put Jev in the right place in a workflow and the benefits show up.