TypeSafe AI Launches Jev, a System One Decision Model

TypeSafe AI, founded by ChatGPT co-inventor Diogo Almeida, released Jev, a new category of AI model designed for fast, structured decisions. Jev does not generate text but outputs calibrated probabilities for classification, routing, and scoring tasks, responding in 70-500ms at $0.042 per million input tokens with free output tokens.
Key insights
Product Overview and Key Specs
Jev is a System One model that inputs unstructured data and outputs type-safe structured results with calibrated probabilities. It does not generate free text, eliminating hallucinations. Response time is 70-500ms, 20-200x faster than frontier LLMs. Input token price is $0.042 per million tokens (about $42 per billion), output tokens are free. The model uses parallel sampling instead of autoregressive generation.
Training Methodology: RLCD
Jev is trained using RLCD (Reinforcement Learning for Calibrated Decisions), a new method developed by TypeSafe CEO Diogo Almeida. He argues that RLHF causes models to hallucinate by rewarding confident-sounding but false outputs. RLCD avoids this by not generating free text; outputs are predefined structures with honest probability estimates. The model never needs to 'make up' an answer.
Use Cases: Smart If-Statements
Jev is designed for embedding AI decisions in software: classification, routing, scoring, extraction, and guardrails. It replaces brittle rule engines or expensive LLM calls in workflows like customer ticket routing, large-scale document classification, real-time recommendations, and LLM output validation. The company calls it 'intelligent if statements'.
Demonstrations: DOOM and Wikiracing
TypeSafe showcased Jev controlling a DOOM AI bot at ~10 queries per second costing ~$7/hour, and playing Wikipedia racing (navigating from one article to another via links). In both cases, Jev outperformed non-reasoning LLMs in speed and step count, demonstrating its suitability for real-time, high-frequency decision tasks.
Philosophical and Economic Rationale
The name Jev references Jevons Paradox: as decision cost drops, total usage increases. TypeSafe bets that making AI decisions cheap enough will unlock massive new use cases. CEO Diogo Almeida's talk 'Why RLCD beats RLHF' laid out the logic: calibrated decision ability allows automation of tasks where the model is 95% correct and can signal when it's in the 5% error zone.
Coding Agent Design Critique
TypeSafe CEO Diogo Almeida published a detailed critique of current coding agents, arguing many design choices are workarounds for KV cache limitations and token pricing. He proposes a 'TypeSafe-centric' design with explicit state and query-aware dynamic context reconstruction, enabling cost-aware routing, sub-agents, and efficient tool use. This vision underpins Jev's architecture.
Availability and Pricing
Jev is currently in early access. Developers can join a waitlist at the TypeSafe website. The company claims it is already profitable and plans to further reduce prices. Output tokens are free, and input pricing is extremely low compared to frontier LLMs.