TypeSafe AI
Sign In
Sept 15, 2026 · TypeSafe News
TypeSafe announces System One models and Jev Read more

Introducing Jev ...... Intelligence beyond chat

The First (Public)
System One Model;
Jev Gives AI The
Properties Of Code

Clock Tool 1.1
Workflow Intelligence vs. Cost
Jev (TypeSafe)193.6x
$0.39
Claude Haiku 4.56.2x
$19.49
Claude Opus 52.1x
$176.05
Claude Sonnet 51.0x
$117.38
Gpt-5.6-Luna6.0x
$3.31
TypeSafe AI
Version 0.01
©2026. All rights reserved.
Made in SF. With Love.
Glider 1.1
Game of Life

We Took The Opposite Research Direction

not chat

RLHF made models superhuman at talking to people. That same loop also made them overconfident, inconsistent, and hard to bury inside software. Chat is a human interface. Automation needs a machine one.

a new model

System One Models are built to be called from code. New architecture, parallel sampler, and a training method we call Reinforcement Learning for Calibrated Decisions (RLCD).

Decisions, not strings

Typed outputs that software can act on.

Calibrated confidence

Every decision includes how sure the model is.

More like code

Reliable, fast, and type-safe.

193.6x Faster, 444.6x Cheaper.

*based on workflows for System One tasks

TypeSafe AI Cost $0.000081
Completed in 0.114s
LLMs Cost $0.013880
Completed in 8.566s

Jev returns typed decisions with calibrated probabilities, so your software can act when confidence is high and escalate when it is not. Combine those decisions in ordinary code.

Jev’s Intelligence Per Dollar Is Literally Off The Charts.

Workflow intelligence vs. cost
Cost per workflow (log) Relative speed Jev · 193.6x · $0.39 Luna · 6.0x Haiku · 6.2x Sonnet · 1.0x Opus · 2.1x
Jev.Cost

$42

Per billion input tokens. Output is free.

vs. Fable 5.1

238x

Lower input price than a frontier chat model.

Come Build With Us

In-person in San Francisco. We want people who care about craft, calibration, and shipping intelligence that software can depend on.

Open roles

TypeSafeAI Blog

Company News

Introducing System One Models & Jev

Read more

Essay

The Bitterest Lesson

Compute drives progress. Progress still has to be pointed at the right task.

Research

Too good to be true, too bad to be useful

Chat models please people. Automation needs decisions machines can trust.

We Give A FAQ

What are System One Models? What is Jev?

System One Models are built for decisions inside software. Jev is the first public one: send structured questions, get typed answers with probabilities and confidence your code can branch on.

Is Jev just a smaller LLM?

No. It does not generate strings. It samples typed fields in parallel, which is why it can be two orders of magnitude faster on System One shaped work.

How is this different from JSON mode?

JSON mode still writes tokens and can break the schema. Jev’s output space is defined up front. Type errors are not a sampling accident; they are not in the space.

How can Jev be so fast and inexpensive?

Parallel sampling, no output tokens to bill, and a stack designed for short structured program state rather than long chat transcripts.

What is Jev good at? Where does it struggle?

Classify, route, score, extract, judge, and guardrail. It is not a writer, coder, or chatbot. If you need a paragraph, use an LLM. If you need a decision, use Jev.

Can Jev still get things wrong?

Yes. Calibration means it tells you when it is unsure. Set thresholds in your own code for when to act and when to ask a person.