Jev

Summary

Jev is a new class of 'System One' frontier model that makes decisions instead of generating text: given unstructured state plus typed questions, it returns typed probabilistic answers (Choice, Score, or Noul/yes-no) with probabilities. It cannot generate text or explain itself. TypeSafe claims 20–200x faster and 40–400x cheaper than frontier LLMs on decision tasks, priced at $0.042/M input tokens with free outputs.

Keywords

decision models · System One · RLCD · agents · structured outputs · TypeSafe

Topics

decision models, System One, RLCD, agents, structured outputs

Research notes

• Discovery: X launch thread by Diogo Almeida (@CompleteSkeptic)
• Method: Trained with RLCD, a new training method developed by Diogo Almeida (ex-OpenAI, InstructGPT/RLHF co-inventor) over two years in stealth; named after William Stanley Jevons (Jevons paradox).
• Key findings: Not an LLM: no text generation, no explanations; all questions in a request evaluated in parallel against shared state; 70–500ms end-to-end latency; 32K tokens per (state + question), 64K per request; $40M seed led by DCVC at ~$200M valuation; available via Vercel AI Gateway as typesafe-ai/jev
• Limitations: Speed/cost claims (193.6x/444.6x in own evals vs GPT-6 Astra/Fable 5.1) are from TypeSafe's own workflow evals; text-only, no images/audio; max 255 options per Choice.
• Details cross-checked across the X launch thread, TypeSafe launch post, and community analyses.