TypeSafe AI launches Jev, a rapid decision-making AI model

TypeSafe AI, founded by a ChatGPT co-inventor, launched Jev, a System One model designed for fast, structured decision-making in software. Jev outputs type-safe results with confidence scores, operates 20-200x faster and 40-400x cheaper than frontier LLMs, and eliminates hallucination by not generating free text.
Key insights
RLCD Training Eliminates Hallucination
TypeSafe AI's Jev uses a novel training method called RLCD (Reinforcement Learning for Calibrated Decisions) instead of RLHF. By not generating free text, Jev cannot hallucinate; it outputs only predefined structured results with epistemically honest confidence scores. This addresses the fundamental issue of RLHF where models learn to feign confidence to please humans.
Performance: 20-200x Faster and 40-400x Cheaper
Jev achieves response times of 70-500ms, compared to 3-329 seconds for frontier LLMs. Input token price is $0.042 per million tokens, with output tokens free. TypeSafe claims Jev is 20-200x faster and 40-400x cheaper than equivalently intelligent models, though extreme multiples represent best-case scenarios. The speed comes from parallel sampling architecture.
Architecture: Parallel Sampling Over Autoregressive Generation
Unlike traditional LLMs that generate tokens sequentially, Jev uses parallel sampling to compute all outputs at once. This architectural shift, analogous to Transformer replacing RNN, enables the dramatic speed improvements. The trade-off is that Jev cannot generate arbitrary text; it outputs only predefined structured decision types.
Use Cases: Intelligent If-Statements for Software
Jev is designed for scenarios requiring a 'smart if-statement' in code: classification, routing, scoring, extraction, and branching. Example applications include real-time game AI (DOOM at 10 queries/second, ~$7/hour), Wikipedia racing, customer service ticket routing, large-scale data processing, and guardrails for LLM outputs. It fills the gap where rule engines are too brittle and LLMs are too slow/expensive.
Product Positioning: Component for Software, Not Chatbot
TypeSafe explicitly positions Jev as a complement to conversational models like ChatGPT and Claude. Jev is not for end users but for developers embedding AI decision nodes in system architectures. The API is not a message stream but structured program state in, typed decisions out. The goal is 'smarter software, not just-in-time software'.
Jevons Paradox: Efficiency Drives Increased Usage
The name 'Jev' references the Jevons Paradox, where increased efficiency leads to greater total consumption. TypeSafe bets that by making AI decisions extremely cheap (output tokens free, input at $0.042/M), the volume of AI usage will explode. The company claims it is already profitable and plans to continue lowering prices.
Availability: Early Access with Waitlist
Jev is currently available in early access. Developers can join the waitlist at TypeSafe's website. The model is already being used in production for various tasks, and the company has demonstrated live demos including DOOM and Wikipedia racing.