Jev: The System 1 Model That 10x's Your Claude Code

In Jay E's RoboNuggets live demo, Jev answers in under a second. The opening claim is clear: built by a team including ChatGPT co-creator Diogo Almeida, this new model runs in a completely different lane from classic chat models. The announcement post reached 38 million views and the message is one line: make decisions in structured form, without generating free-form text.

Price and Speed: Why So Cheap and Fast?

The numbers are aggressive. TypeSafe's pricing shows Jev at about 24x cheaper than Haiku and 230x cheaper than Fable 5.1. Input costs $0.04 per million tokens, output is free. The headline claims are 20 to 200x speed and 40 to 400x lower cost; in the controlled workflow measurements on the TypeSafe blog the ceiling is given as 193.6x faster and 444.6x cheaper. So not your everyday average, but a lab ceiling under specific workflows.

The secret is constraint. Jev answers in only three shapes: binary true/false, pick from a list, and a 0-to-10 scale. It looks like a heavy limitation at first, but Jay E's point is that the narrow output stops the model from fabricating text; instead it returns a well-calibrated decision with probabilities. With all that generative variance removed, latency and cost collapse.

TypeSafe frames the split with Daniel Kahneman's Thinking, Fast and Slow. Jev is the first System 1 model; Fable and Astra-like large language models are System 2. System 2 gains flexibility by emitting token by token but gets slow. System 1 is fast, intuitive, schema-guaranteed and does not hallucinate. The best architecture is to pair them: Jev for fast classification, Claude for nuanced generation.

Access is now open. There was a waitlist during recording, but TypeSafe opened early access to everyone by the time the video went live. You can get a key directly from TypeSafe or connect in one command via OpenRouter. Jay E used OpenRouter in his own setup and showed a single-command setup instruction on screen; a free PDF with all commands sits in the description.

Level 1: Speeding Up the Agent OS

Level one is making your existing agent setup faster and cheaper. First trick is model routing. Using the most expensive model for everything makes no sense; Opus and Fable are pricey while Haiku and Sonnet handle many tasks fine. The problem was that picking the right model fell to you. Jev automates that choice. In Jay E's 12-prompt comparison, Jev-guided routing saved 70% because 9 tasks did not need the strongest model at all.

In practice he suggests a skill command to toggle it: type /Jev on and every task in that session gets routed by Jev to the best-fit model. In the example, finding the file path of the Jev prompt was routed to Haiku; without Jev the session default would have been Opus 5. Token economy is protected by a one-line router.

Second trick is skill discovery. Jay E's workspace holds 145 skills. In a 14-test benchmark, Jev found the correct skill in about 5 seconds, while the default Opus 5 approach took 30 seconds. Logic is simple: input is your task, options are your skill list, Jev pulls the right skill instantly for Claude to load. A one-command instruction can be copied to try it yourself.

Levels 2 and 3: From Business Automation to New Apps

Level two is business automation. The demo classifies 800 emails with a single business question: which ones are real prospects to contact? Jev, Haiku and Fable run side by side; on play, Jev tags the whole list as interested / not a prospect / not interested in under a second. The others are visibly slower and more expensive. The pattern is clear: high-volume data plus a single business question is a perfect fit for Jev.

Jay E generalizes the pattern. Invoice fraud detection is a huge industry; spam filters, community management, high-volume refund requests and churn prediction for subscription businesses fit the same template. The question to ask in your company is: what do you receive in bulk that needs classifying? Put Jev there and automation drops to a few text commands.

Level three is apps now possible because the cost barrier fell. In Jay E's Rubrik system holding hundreds of images and videos, filename search finds only files with that word in the name while Jev-powered semantic search returns by meaning and typing Claude surfaces genuinely Claude-themed visuals. In the same logic Kitsi's Chrome extension Unclutter classifies every page element in the background with Jev and cleans ads and cookie banners in one click.