Introducing System One Models & Jev: A New Frontier for AI-Powered Automation

TypeSafe AI has spent two years solving a deceptively simple problem: if LLMs are superhuman at chat, why hasn’t automation followed? Their answer is a new model class called System One Models and a first release named Jev—built not to generate text, but to make fast, structured, type-safe decisions software can use directly.

The Gap Between Chat and Automation

Existing LLMs optimize for human preference. They generate strings—flexible, powerful, and prone to hallucinations, type errors, and unpredictable outputs. Before software can act on those outputs, you must parse, validate, and hope the model stayed on track.

Jev takes the opposite approach. You define possible outputs in advance. The model returns typed, structured values with calibrated confidence scores—and mathematically cannot produce a type error or hallucinate.

How System One Models Differ

TypeSafe trained Jev using Reinforcement Learning for Calibrated Decisions (RLCD), replacing the human-preference feedback loops of standard LLMs. Three architectural differences drive the results:

  • Parallel sampling: Jev generates all outputs in a single query instead of one token at a time, enabling 70–500ms end-to-end response times versus 3–329 seconds for frontier LLMs.
  • Structured outputs only: Inputs are unstructured data; outputs are type-safe values with confidence scores. No string generation means no hallucinations.
  • Calibrated confidence: Every output includes a probability score. Higher confidence correlates directly with higher accuracy, making Jev composable into automated pipelines without human oversight.

Performance

On TypeSafe’s workflow evaluations—designed to measure AI performance inside real production code—Jev achieves comparable intelligence to top frontier models while running up to 200x faster and 444x cheaper ($0.042/MTok input, free output tokens).

The evaluations test four production-grade workflows: classify, route, score, and extract tasks where hand-written logic breaks down. Jev dominates the Pareto frontier across both speed and accuracy dimensions.

What You Can Build

System One Models fit anywhere you need a reliable decision inside code:

  • Smart conditional logic: Replace brittle if-statements with fuzzy classification and routing
  • Real-time applications: Sub-100ms latency makes AI viable where UX demands it
  • Large-scale data processing: Map petabytes of unstructured data into typed features
  • LLM guardrails: Score, judge, and detect jailbreaks in other models’ outputs

TypeSafe has demonstrated Jev playing Doom reactively at 10 queries per second and winning Wikiracing challenges against frontier LLMs—tasks that compound the benefits of speed, accuracy, and zero hallucinations.

Get Started

Jev is available today in early access. TypeSafe is pulling developers off the waitlist now and wants to know which decisions you need to automate. Start at typesafe.ai and review the docs to define your first structured output schema.

The cost of intelligence is dropping fast. Every order of magnitude unlocks new use cases—and Jev is designed to accelerate that curve.