TypeSafe AI Launches Jev, a System One Decision Model

TypeSafe AI, founded by ChatGPT co-author Diogo Almeida, released Jev, a new category of AI model called System One. Jev generates no free text; instead it outputs type-safe structured results with calibrated probabilities. Designed for quick decisions in code (classification, routing, extraction), Jev is 20-200x faster and 40-400x cheaper than frontier LLMs for System One tasks, with response times of 70-500ms and free output tokens.
Key insights
Architecture: No Text Generation, Parallel Sampling, Type Safety
Jev adopts parallel sampling where all outputs are computed at once, unlike serial token generation in LLMs. It does not generate free-form text; output structure is predefined, making type errors mathematically impossible and eliminating hallucination. Each output comes with an epistemologically honest probability. This is based on RLCD (Reinforcement Learning for Calibrated Decisions), a new training method developed by TypeSafe.
Performance and Pricing
Response time 70-500ms vs. frontier LLM's 3-329 seconds. Input token price $0.042 per million tokens; output tokens free. TypeSafe claims 20-200x faster and 40-400x cheaper at equivalent intelligence on System One tasks. Workflow evals used GPT-6 Astra and Anthropic Fable 5.1 as references; Jev may be underestimated in relative performance. However, extreme speed multipliers (e.g., 193x, 444x) represent best-case scenarios.
Target Use Cases: Intelligent If-Statements
Jev is designed as a component in code that needs AI judgment but not expression. Use cases include: intelligent branching in workflows (routing tickets, sentiment detection), large-scale data processing (classification, labeling, extraction), real-time applications (game AI, dynamic pricing), and guardrails (checking LLM outputs for safety, format, jailbreak). The company calls these 'smart if statements'—places where rule engines are too brittle and LLMs are too slow/costly.
Community Demonstrations and Experiments
Several demos showcase Jev's capabilities: real-time DOOM AI control (~10 calls/sec, $7/hour), Wikipedia racing game (fewer steps than non-reasoning LLMs), semantic search (Ctrl+F with meaning, e.g., finding 'remote work' in PostgreSQL), and a game building by the Manus team using Jev for rapid decisions in browser and 3D rendering. These illustrate how fast, cheap decision models enable new interactions.
Product Positioning and Pricing Philosophy
TypeSafe's CEO explained that they are betting on task space expansion rather than unit intelligence price war. The name 'Jev' references the Jevons paradox: as efficiency of a resource improves, total consumption increases. Free output tokens reflect that decision cost is so low it's not worth billing. The company already claims profitability with plans to keep lowering prices. This positions Jev not as a competitor to LLMs but as a complement for deterministic decision nodes.