Last updated: 2026-09-20 · Sources: typesafe.ai, provider listings, independent write-ups (linked inline)

What Is Jev? TypeSafe's System One Model Explained

Jev is the first of a new model class — System One Models — announced by TypeSafe AI on September 15, 2026. Instead of generating text, it answers typed questions about your input and returns structured decisions with calibrated probabilities. This guide explains how it works, what it can and cannot do, and which claims you should verify yourself.

TL;DR

• What: a model that returns typed decisions (a choice, a score, or a true/false probability) with confidence estimates — never free-form text.
• Why it matters: ~193.6x faster and ~444.6x cheaper than frontier LLMs on structured-decision workflows (TypeSafe's own numbers), because there is no sequential token generation.
• Price: $0.042 per 1M input tokens; output is free.
• Status: early access — official waitlist, or instant via Vercel / Cloudflare / OpenRouter.
• Catch: benchmarks are vendor-run and not independently reproduced; a valid answer can still be wrong with high confidence.

The problem with chat models

RLHF-era LLMs are optimized to produce words people like. That makes them superhuman at instruction following — and unreliable as software components: they drop modes, express overconfidence, hallucinate free-form text, and wrap every answer in prose. If what your application actually needs is “which of these 4 categories is this ticket?” or “is this claim supported by this source?”, a chat model is an expensive, noisy way to get a one-bit answer.

TypeSafe's bet is that a large share of what people currently ask LLMs to do is structured decision-making dressed up as chat — and that this share deserves its own model class, natively consumed by machines.

How Jev works

Input: state + typed questions

You send one HTTP request containing a state (a block of text to judge — a ticket, a paragraph, a log line) and one or more questions, each with a fixed answer type. The response returns every answer with a probability, in a single pass.

{
  "state": "Hi, I've been trying to connect my Stripe account
    for 3 days and it keeps failing. I'm losing sales.",
  "questions": {
    "urgency":   { "type": "noul",   "statement": "This message expresses urgency." },
    "category":  { "type": "choice", "options": ["billing", "technical", "sales", "spam"] },
    "frustration":{ "type": "score",  "scale": ["low", "medium", "high", "critical"] }
  }
}

Training: RLCD, not RLHF

TypeSafe built Jev with a new architecture, a new sampler, and a training algorithm called Reinforcement Learning for Calibrated Decisions (RLCD). The optimization target is calibration — a 0.9 should actually mean 90% — rather than human preference on generated prose. This is also why the model cannot produce invalid answers: the output shape is fixed at the architecture level, not by prompt engineering.

Output: decisions + confidence thresholds

Your code owns the policy: auto-execute above 0.9, escalate to a human between 0.6 and 0.9, fall back to a frontier LLM below that. TypeSafe's framing: “Combine those decisions in code to build larger workflows, with control over how the intelligence is used.”

The three primitives

All three can share one request — a common pattern is fanning out every question you might need (speculative fan-out) and letting code pick the relevant answers, because adding questions barely moves the bill when input is the only thing priced.

Benchmarks — and their caveats

TypeSafe ran a custom 4-workflow evaluation (security incident response, agent-trace observability, invoice processing, customer service):

Pricing in one table

Full provider-by-provider rates and an interactive cost calculator live on the pricing page.

Limitations

• No text generation. Jev cannot write an essay, an email, or a chatbot reply. Any human-facing prose still needs an LLM.
• Text only. No image or audio input yet.
• 64K context (32K on Cloudflare). Long documents need chunking or summarizing first.
• No arithmetic, no dates. Don't ask it to count, sum, or compute date differences — do that in code.
• Context-sensitive. Strip tracking headers, HTML boilerplate, and unrelated metadata from the state before sending.
• Confidently wrong is possible. Calibration helps, but a wrong answer can carry high confidence. Regression-test probabilities against historical data before automating decisions.
• Early-access economics. Pricing and rate limits may change; TypeSafe says it can't prove the price isn't subsidized.

The right mental model

Jev isn't a cheaper LLM. It's a typed function call that happens to be intelligent: it returns a value your code can use directly, with a confidence number attached. Once you have that, a lot of code that exists only to survive and validate string output stops needing to exist at all. Within days of launch, the open-source openjev project reproduced the interface pattern on a small open model — a sign the primitive may outlive any single vendor.

How to try it

Three routes, from zero-code to full API access, are covered step by step in our Get Access guide — including the no-waitlist paths via Vercel, Cloudflare, and OpenRouter. For when to use Jev versus a frontier model, see Jev vs LLMs; for concrete workflows, see use cases.