Jev found a Zurich to London flight in 7 seconds, and most developers still don’t understand how.
Jev physically doesn’t write text. It can’t form a sentence, explain a decision, or generate code. It gets the current state and returns exactly one of three things: a choice from a list, a score, or a number between 0 and 1.
That’s the whole secret behind the flight demo. The page shows the available actions, Jev chooses the next click, the page changes, and Jev chooses again. No text between the steps - just a chain of decisions, one after another, until the options appear.
Each decision costs $0.042 per million input tokens. Not because the model was trained to be cheaper, but because generating text was never necessary in the first place.
The same approach was applied not to flights, but to Claude’s own memory: instead of summarizing the conversation history again, Jev simply removed everything that no longer mattered. A session of nearly one million tokens collapsed to 86,000. In one second.
Developers have been paying for generation for years where all they needed was a simple “yes,” “no,” or a number from a list. An expensive model was writing a paragraph to answer a question that had exactly two possible answers.