Jev — TypeSafe AI's First System One Model
A fast intro to the new decision-only model class — what it is, how to call it, and how it might reshape agent architecture.
- Buy once, yours for good
- 14 lessons across 5 sections
- Progress tracking across devices
What you'll learn
- Explain what a "System One model" is and how it differs from an LLM
- Read a Jev API call and know what Choice, Score and Noul each return
- Recognise the workloads where a decision-only model beats a general LLM
- Weigh TypeSafe's own benchmark claims against how they were actually measured
- Sketch an agent architecture that uses Jev as a fast front door to a slower LLM
- Know what's still unverified about Jev, and what to watch for as it leaves early access
Course content
5 sections · 14 lessons
Meeting JevWhat shipped, who built it, and why it isn't trying to be a chatbot.2 lessons
- The thirty-second versionFree preview2m
Read this lesson
TypeSafe AI is a San Francisco startup founded in 2024 by Diogo Almeida — who spent four years at OpenAI working on RLHF, InstructGPT, ChatGPT and GPT-4 before leaving in 2024 — along with Erik Gafni and Sasha Sheng. The company came out of stealth on 15 September 2026 with a $40M seed round led by DCVC, a reported $200M valuation, and one public product: Jev.
Jev doesn't write text. You send it a state — some unstructured context — and a set of typed questions. It answers all of them in a single pass, with a probability and a confidence score attached to each answer, in 70–500 milliseconds.
- System One vs System TwoFree preview5m
Read this lesson
The name is the pitch. Kahneman's "System 1 / System 2" split — fast, automatic, intuitive judgement versus slow, deliberate reasoning — has become a common shorthand in AI for two kinds of work software needs: quick classification and routing decisions, and open-ended reasoning and generation. Frontier LLMs are built and priced for the second kind. TypeSafe's bet is that the first kind deserves a purpose-built model instead of a general one running in an expensive, overqualified mode.
Concretely, Jev differs from an LLM in three ways:
- Non-autoregressive. It doesn't generate token by token — all outputs for a request are produced together, which is most of where the latency and cost advantage comes from.
- Schema-constrained. Every answer is typed and bounded to the options you defined. It cannot emit a value outside your schema — TypeSafe frames this as eliminating a category of hallucination by construction, not by prompting.
- Trained differently. TypeSafe describes a method called RLCD (Reinforcement Learning for Calibrated Decisions), trained entirely on synthetic data, optimising probabilities against outcomes rather than human preference ratings the way RLHF does.
One honest gap: TypeSafe has not published a technical paper, architecture details, or weights. It says Jev is transformer-based; outside observers have speculated it's built on top of an existing open-weight LLM. Treat the architecture as a vendor description, not a verified fact.
Calling JevThe actual shape of a request, with the primitives and a worked example.3 lessons
- The three primitives6m
- A worked example: triaging a support ticket8m
- The Python shape4m
What it costs, and how fast it actually isThe headline numbers, and then a closer look at how they were produced.2 lessons
- The headline numbersFree preview4m
Read this lesson
TypeSafe's pricing: $0.042 per million input tokens, output priced at $0, and end-to-end latency of 70–500ms. The company's speed and cost claims vary by where you read them — the launch page headlines 193.6x faster / 444.6x cheaper, the blog body says 40–200x faster for System One-shaped queries, and the founder's own thread says 20–200x faster, 40–400x cheaper. All three are describing the same launch.
- Reading the benchmark carefully6m
Where you'd actually use thisTypeSafe's own framing, plus some speculation about adjacent uses.3 lessons
- Real-time and embedded decisions5m
- Agent guardrails and routing — speculative6m
- A sketch: Jev in front of a coding agent — speculative6m
Speculating on the bigger pictureTypeSafe's own bet, and the honest limits of what's known nine days in.4 lessons
- The Jevons paradox bet6m
- What's still unverified5m
- What to watch next3m
- Sources3m
Requirements
- Comfortable reading a JSON API response
- Helpful but not required — some sense of how LLMs and agents work
Description
On 15 September 2026, a two-year-old stealth startup called TypeSafe AI shipped something that is not a chatbot and does not generate text: Jev, the first model in a category TypeSafe calls "System One models" — fast, typed, schema-constrained decisions instead of prose.
This is a short course, on purpose. Jev is days old, gated behind a waitlist, and mostly unverified outside its own vendor — so the goal here is not a verdict. It's enough of a tour that you can read the announcement, try the API, and form your own opinion about where this does and doesn't matter.