What is Jev?

Jev is the first System One Model released by TypeSafe AI — a new class of frontier models built specifically for fast, structured decisions that software can use directly. Unlike traditional LLMs optimized for chat and text generation, Jev is designed from the ground up for automation: it takes unstructured program state as input and returns typed, calibrated probabilistic decisions as output.

"Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." — Diogo Almeida, Founder, TypeSafe AI

The Core Innovation: System One vs. System Two

Inspired by Daniel Kahneman's Thinking, Fast and Slow, TypeSafe distinguishes between:

AspectTraditional LLMs (System Two)Jev (System One)
TrainingRLHF / RLVR (human preferences)RLCD — Reinforcement Learning for Calibrated Decisions
Optimization TargetHuman-preferred chat responsesCalibrated decisions with epistemically honest probabilities
InputSequential messages (chat)Structured program state
OutputStrings (text, code, hallucinations)Type-safe structured values (Choice, Score, Boolean)
SamplingSequential (token-by-token)Parallel (all outputs in one query)
Speed3–329 seconds70–500ms (40–200x faster)
Cost$0.20–$10/M input tokens; output ~5x more$0.042/M input tokens; output FREE
HallucinationInherent riskMathematically impossible (schema-guaranteed)
ConfidenceOverconfident, inconsistentCalibrated — higher c higher accuracy

Why Jev Changes Everything

1. Zero Hallucination, Guaranteed Type Safety

Jev cannot hallucinate a format or emit an invalid type. The output schema is defined in advance — every answer conforms to your declared types. This is "table stakes for automation" — a hallucinated tool call is inconvenient in a chatbot, but a deal-breaker in production systems with latency guarantees or deep dependency chains.

2. Calibrated Probabilities, Not Guesses

Every Jev output includes calibrated confidence scores. If Jev says 95% confidence, it's right 95% of the time. Traditional LLMs are notoriously overconfident and inconsistent — they can't tell you when they're in the 5% failure case, making them unreliable for automation.

3. Parallel Sampling = Insane Speed

Jev evaluates all declared questions in a single parallel pass instead of autoregressive token generation. This enables:

  • 70–500ms end-to-end latency (vs. seconds for LLMs)
  • Real-time AI in UX-critical paths
  • Map-reduce over petabytes of data
  • 10+ queries/second in game loops (see the Doom demo)

4. Economics That Enable New Use Cases

  • Input tokens: $0.042/M ($42 per billion)
  • Output tokens: FREE (too cheap to meter)
  • 193.6x faster, 444.6x cheaper on workflow evaluations vs. frontier LLMs

This follows the Jevons Paradox (namesake William Stanley Jevons): every order-of-magnitude drop in intelligence cost unlocks orders of magnitude more use cases.


Ideal Use Cases for Jev

Jev shines where you need structured decisions at speed and scale:

CategoryExamples
Smart Routing / ClassificationSupport ticket triage, intent classification, spam detection
Risk Scoring & GuardrailsFraud scoring, content moderation, jailbreak detection, refund approval
Agent Control LoopsChoose next tool, decide continue/retry/stop, subagent selection
Real-Time ApplicationsGame AI (Doom bot), browser agents, live recommendations
Data Processing at ScaleExtract fields from millions of documents, feature engineering
Verification & EvaluationScore LLM outputs, verify reasoning traces, enforce policies

Reference: TypeSafe AI Official Announcement

📖 Primary Source: Introducing System One Models & Jev — TypeSafe AI Blog

Published September 15, 2026 by Diogo Almeida (founder, former OpenAI). The post includes:

  • Side-by-side demos comparing Jev vs. GPT-5.6 Terra
  • Workflow evaluations showing Pareto-frontier performance
  • Technical deep-dive on RLCD training methodology
  • Hallucination/type-safety guarantees
  • Fun demos: Doom bot and Wikiracing agent
  • FAQ addressing "Is Jev just a smaller LLM?" and "How is this possible?"

Best GitHub Repositories to Get Started

🏠 Official TypeSafe AI Organization

RepoDescription
typesafe-ai/system-one-adapter-pythonOfficial Python wrapper to constrain LLMs to System One structured outputs
typesafe-ai/typesafe-ai.github.ioTypeSafe AI website & docs source

🚀 Community Quickstarts & Playgrounds

RepoDescriptionBest For
codaaiteam/jev-aiQuickstart & FAQ — 30-second intro, free browser playground at jevtypesafeai.comBeginners, immediate hands-on
codaaiteam/jev-typesafe-aiDeveloper notes & examples — Practical patterns, live playground running real /v1/systemone callsDevelopers building integrations
TypeSafeAI/typesafe-playgroundInteractive playground — Edit classification experiments, A/B inputs, routing, extraction, code-policy decisions, gamesExperimentation, learning by doing

🔧 Integrations & Tooling

RepoDescription
Vercel AI SDKOfficial Jev support via experimental_evaluate API (AI SDK 7.0.105+) — model: 'typesafe-ai/jev'
awesome-jevCurated list of Jev tools, integrations, examples (community-maintained)
jev-mcpModel Context Protocol server for Jev

Quick Start: Your First Jev Call (TypeScript)

pnpm add ai@latest
import { experimental_evaluate as evaluate } from 'ai';

const result = await evaluate({
  model: 'typesafe-ai/jev',
  state: 'The support agent issued a full refund to the customer.',
  questions: {
    refunded: {
      type: 'boolean',
      instructions: 'Was a refund issued?',
    },
    priority: {
      type: 'choice',
      options: ['low', 'medium', 'high', 'critical'],
      instructions: 'What is the urgency level?',
    },
    riskScore: {
      type: 'score',
      rubric: ['safe', 'review', 'block'],
      instructions: 'Risk level of this action',
    },
  },
  providerOptions: {
    gateway: { zeroDataRetention: true },
  },
});

console.log(result.answers.refunded);      // { value: true, probability: 0.99 }
console.log(result.answers.priority);      // { value: 'high', probability: 0.92 }
console.log(result.answers.riskScore);     // { value: 'review', probability: 0.78 }

// Calibrated confidence metadata
console.log(result.providerMetadata?.typesafe?.confidence);

What's Next for Jev?

TypeSafe AI is still in early access — they're actively onboarding developers from the waitlist. Key areas to watch:

  1. Expanded cardinality support (currently up to 255 options per Choice)
  2. Multi-modal inputs (images, structured logs, traces)
  3. Fine-tuning / customization for domain-specific workflows
  4. Lower latency tiers for ultra-real-time use cases
  5. Open benchmarks and community evaluation standards

Final Thoughts

Jev represents a fundamental shift in how we integrate AI into software:

Old way: Prompt an LLM → Parse text → Validate → Hope it doesn't hallucinate → Retry on failure

Jev way: Define typed questions → Call Jev → Get calibrated, type-safe decisions instantly → Branch with confidence

If you're building AI-powered workflows, agents, real-time apps, or data pipelines — Jev is worth a serious look. The economics and latency finally make "AI as a reliable function call" viable.


Resources


Have you tried Jev? Building something with System One Models? Share your experience in the comments or reach out on X.