When TypeSafe AI came out of stealth in September 2026 with founder Diogo Almeida, a name attached to both ChatGPT and RLHF, the framing was almost contrarian: not a smarter chatbot, but a model that refuses to write free text at all. Jev, their first "System One model," only answers three kinds of questions: pick one option from a list (Choice), rate something on a scale (Score), or estimate the probability that a statement is true (Noul). No prose, no reasoning trace, just a fast, calibrated, typed answer. The claims, up to 200x faster and 400x cheaper than a frontier LLM for these tasks, sounded like marketing until you look at what people are actually shipping with it.
Browser agents that stop thinking about where to click.Use Jev to look at a snapshot of a webpage's DOM and pick the next element to interact with directly, instead of asking a full LLM to reason through "which button is Submit." That single substitution takes the decision from a multi-second reasoning call down to roughly 300 milliseconds, which is the difference between a browser agent that feels responsive and one that feels like it's thinking too hard about a task a human does without looking.
Game AI with real reflexes. There's a small but telling cluster of Jev-powered bots for Tetris, Pac-Man, Snake, gomoku, and even Doom. None of them ask the model to "reason about strategy." They ask it to score the handful of legal next moves given the current board state, every frame, cheaply enough that latency isn't the bottleneck. It's a nice demonstration that a lot of what looks like "AI gameplay" is really just fast, well-calibrated classification repeated very quickly.
Guardrails that can afford to check everything. Content moderation and prompt-injection detection are natural fits for a Noul-style yes/no probability: is this message trying to manipulate the system, does this retrieved passage contradict the source document, is this output safe to show a user. Because each check costs a fraction of a cent, teams are running Jev as a screen on every message in and out of an LLM pipeline, rather than sampling a subset the way you might with a heavier model.
Coding-agent context management. One of the more inventive uses so far is save-token-jev, which tackles a real pain point in tools like Claude Code and Codex: context windows filling up with old tool calls. Instead of burning an LLM call to summarize (and risk losing detail), it asks Jev a Noul-style question per past tool call, is this still relevant, and keeps or drops accordingly, with user and assistant text always left untouched. It's a clean example of using Jev not as a feature, but as plumbing.
Decisions where milliseconds are money. On the more extreme end, jev-trader makes a single buy/sell/hold call per new blockchain block based on order-book state. Whether or not you trust an AI with that job, it's a useful stress test of the latency claim. This is a domain where "fast enough" has a literal price attached.
What ties all of this together is that none of these tasks needed a model that can write. They needed a model that can decide, quickly and cheaply, thousands of times a day. Whether Jev's calibration claims hold up under independent scrutiny is still an open question, and it's worth treating the 200x/400x numbers as ceiling cases rather than typical results. But as a signal for where AI infrastructure is heading, not bigger chat models, but cheap, fast judgment calls wired directly into software, it's a genuinely interesting one to watch.
Top comments (0)