What is Jev?
Jev is a new class of frontier AI model called a System One model, developed by TypeSafe AI. Instead of generating slow, conversational text, Jev evaluates an input state and returns typed, calibrated decisions that software can immediately execute.
1. System 1 vs. System 2 Thinking
The name draws from Daniel Kahneman's cognitive theory. Human cognition operates in two modes:
System 2 (Deliberative)
Autoregressive token-by-token generation. Slow (1-10 seconds), expensive ($2-$20/M tokens), non-deterministic strings that require JSON parsing and regex verification.
System 1 (Intuitive & Fast)
Non-autoregressive parallel sampling. Sub-100ms response, $0.042/M input tokens, free output tokens, and guaranteed valid typed schemas.
2. Core Primitives: The Three Decision Types
Discrete Classification
Picks 1 option from a list of pre-defined choices with calibrated confidence probabilities.
Calibrated Scoring
Returns an exact numeric judgment from 0 to 100 with probability distribution and uncertainty bands.
Binary Confirmation
A high-confidence boolean gate (Yes/No) with fallback thresholds for human intervention.
3. Speed and Economics
Because Jev evaluates all output questions in parallel across a single forward pass rather than generating words one after another:
- Latency: 70ms to 250ms end-to-end, making it suitable for 60Hz robotic loops, game physics, and browser DOM interactions.
- Pricing: $0.042 per million input tokens, with output decisions completely free of charge.
- Reliability: 100% schema guarantee. Impossible to produce unparseable markdown codeblocks.