Models have been superhuman at chat for years, so where is all the automation?
That question is why this lab exists. Chat made language models useful to people. It did not make them a primitive that software can call the way it calls a function.
After two years in stealth, TypeSafe is releasing the first System One Model: a frontier model built to make fast, structured decisions that code can use directly.
The stack is new end to end. New architecture. A parallel sampler. And a training method we call Reinforcement Learning for Calibrated Decisions (RLCD).
Our first public model is Jev, in early access today. On System One work it sits near frontier intelligence, while running about two orders of magnitude faster. It does not write strings. It cannot hallucinate a type. Unstructured state in. Typed probabilistic decisions out.
Extraordinary claims need evidence. The rest of this note is the receipts.
RLHF / RLVR
Reinforcement Learning for Calibrated Decisions (RLCD)
Human preference: writeups and chat that raters like.
Verifiable rewards: outputs a program can check.
Calibrated decisions: honest probabilities on System One tasks.
Unstructured data, with an emphasis on sequential messages.
Unstructured data, with an emphasis on structured program state.
Strings. Flexible, and therefore also able to refuse, wander, or break a schema. Software has to parse and hope.
Type-safe values. The shape is defined in advance. Every answer ships with a probability and a confidence score.
Sequential. One token, then the next.
Parallel. All fields in one shot. Hardware-aware and cheap.
Input: $0.20 to $10 / MTok.
Output: about 5x the input price.
Input: $0.042 / MTok ($42 per billion tokens).
Output: free. Too cheap to meter.
End-to-end: 3 to 329 seconds for frontier models. Fine for a person. A bottleneck in a loop.
End-to-end: 70ms–500ms. Roughly 40x–200x faster on System One shaped queries.
Asked for a number, models still overclaim. If they are right 95% of the time and never flag the 5%, you cannot automate the task.
Every output states uncertainty. Higher confidence tracks higher accuracy. Similar inputs get similar answers.
Human-in-the-loop work: chatbots, copilots, coding agents.
Verifiable problems: proofs, kernels, anything you can check for free.
Demos that only have to work sometimes.
Smart if-statements: classify, route, score, extract, branch.
Map-reduce over large corpora.
Real-time product surfaces around 100ms.
Verify, judge, and guardrail other models.
We like skeptics. We are skeptics.
Easy to check: latency from a West Coast laptop, public pricing, and zero type errors — the last one is a property of the output space, not a leaderboard score.
The bolder claims get more nuance below.
Jev fills every field at once. An LLM writes a string, token by token. Giving up generation is how the rest of the superpowers show up.
The account is… maybe at risk? Let me think about usage, tickets, and the last invoice before I…
We score models inside a fixed compute graph, not a free-form prompt. The graph stays still. The reference is the average of the strongest external models. Jev sits on the Pareto frontier for almost two orders of magnitude.
Real workflows ask many independent questions and branch on probabilities, not slogans. That is the work automation actually is.
A bad tool call is annoying in a chat agent. It is a deal-breaker in a latency budget or five layers down a graph. Schema match for Jev is guaranteed, so the type-error rate is zero by construction.
Two favorites from the team.
A bot reading structured game state, ten calls a second, about $7 an hour. Not the strongest Doom player possible — the point was following instructions against changing representations of the world, in real time.
Start on one Wikipedia page, reach another using only the links you find. Cardinality in the hundreds or thousands. Speed matters; so does never inventing a link that is not there. High-cardinality choices above 255 use a two-stage score-then-pick.
Jev is early. Early access is opening as fast as we can pull people off the waitlist. Tell us which decisions you need to automate, where it works, and where it fails.
We started TypeSafe because intelligence needed an interface software could depend on. We want the next uses to show up in other people’s code, not in a demo reel.
System One is from Kahneman: fast, intuitive judgment versus slow deliberation. The name has also meant “error-prone.” We think this class can be made more reliable than the alternative, and we will write more on why.
Jev is for William Stanley Jevons. When steam engines got more efficient, coal use went up. We expect cheap intelligence to work the same way: every order-of-magnitude drop in cost unlocks more work, not less.
RLHF optimizes for what a person likes to read. RLVR optimizes for what a checker accepts. Neither asks the model to be honest about probability on a decision software will actually take.
Routing, moderation, extraction, scoring, review, and any place you wish a brittle if-statement could see. Not novels, not repos, not a friend.
It is a different output interface and a different sampler. Size is not the point. Structure is.
Public chat and coding benches measure string generation. We publish workflow evals instead, because that is the job.
Stop generating text. Sample the decision. Bill the input. That is most of the trick, plus a stack that never pretended to be a chatbot.
Early access is opening in waves. Leave an email and we will move you up as seats open.