TypSafe AI Launches Jev, a System One Model without Text Generation

TypeSafe AI, founded by a ChatGPT co-author, released Jev, a new category of AI model called System One that outputs structured decisions with calibrated probabilities instead of generating text. Jev is designed for fast, low-cost decision-making in software, with response times of 70-500 ms and input pricing at $0.042 per million tokens, with free output tokens.
Key insights
Jev's Core Design: No Text Generation, Only Structured Decisions
Jev completely abandons text generation to avoid hallucination. It outputs predefined structured results with attached confidence scores, making type errors mathematically impossible. This design stems from the insight that RLHF encourages models to feign confidence, while Jev's RLCD training method produces calibrated decisions. The model is optimized for tasks like classification, routing, scoring, and extraction, acting as an 'intelligent if statement' in code.
Performance and Cost Advantages
Jev achieves 70-500 ms response times, 20-200x faster than leading LLMs, and is 40-400x cheaper, with input at $0.042 per million tokens and free output. This speed comes from parallel sampling instead of autoregressive token generation. The trade-off is that Jev cannot generate free text, but it excels in scenarios requiring fast, reliable, structured decisions.
Target Use Cases and Demonstrations
Jev is designed for developers embedding AI decisions in code: intelligent branching in workflows, large-scale data processing, real-time applications, and guardrails for LLM outputs. Demonstrations include controlling a DOOM bot at 10 queries per second ($7/hour) and Wikipedia racing. TypeSafe also showed integrations like semantic Ctrl+F in web pages and a PostgreSQL function for semantic search.