TypeSafe AI launches Jev, a System One decision model

TypeSafe AI, founded by a ChatGPT co-author, released Jev, a new category of AI model called System One. Jev outputs structured, type-safe decisions with calibrated probabilities in 70-500ms, costing $0.042 per million input tokens with free output tokens. It is designed for fast, reliable decision-making in software systems.
Key insights
Jev abandons text generation to eliminate hallucinations
Jev does not generate free text; it outputs predefined structured results with calibrated confidence scores. Type errors are mathematically impossible because output structure is enforced at the architecture level. This approach directly addresses the hallucination problem inherent in RLHF-trained LLMs, which learn to be overconfident and fabricate answers to please humans.
Jev is 20-200x faster and 40-400x cheaper than frontier LLMs
Jev achieves 70-500ms response time compared to 3-329 seconds for frontier LLMs. Input tokens cost $0.042 per million, output tokens are free. TypeSafe claims Jev is 20-200x faster and 40-400x cheaper than models of equivalent intelligence. Speed comes from parallel sampling instead of sequential token generation.
Jev targets 'intelligent if statements' in software systems
Jev is designed for tasks like classification, routing, scoring, extraction, and branching in code. It serves as a fast, reliable AI decision node where rule engines are too brittle and LLMs are too slow or expensive. Use cases include customer support routing, large-scale data processing, real-time applications, and guardrails for LLM outputs.
New three-layer architecture for AI applications
Jev introduces a three-layer architecture: LLM layer for open-ended work (coding, research), Jev layer for fuzzy but bounded semantic decisions (routing, risk gating), and deterministic code layer for hard constraints (file permissions, spending limits). This separation allows each layer to be optimized for its role.
Jev enables new applications like real-time game AI and semantic search
Demonstrations include controlling a DOOM AI bot at 10 calls per second ($7/hour), Wikipedia racing, and a 3D rendering game. Third-party demos show semantic Ctrl+F in browsers, semantic retrieval in PostgreSQL, and bug detection from user event logs. These showcase Jev's ability to make fast decisions in interactive systems.