TypeSafe AI introduces Jev and its System One Models

Summary

TypeSafe AI has introduced Jev, its first public System One Model, as an alternative to language models optimized primarily for human-facing chat. The company argues that reinforcement learning from human feedback can contribute to mode dropping, overconfidence, and unreliable decisions that require human review. System One Models are designed to produce typed decisions that software can act on, rather than only generating strings for people. TypeSafe says its approach combines a new architecture, a new sampler, and a training method called Reinforcement Learning for Calibrated Decisions, or RLCD. Each Jev decision includes probabilities and a confidence estimate, allowing developers to set thresholds for autonomous action and escalation to review, then combine decisions into larger workflows. The company presents Jev as reliable, fast, self-consistent, and type-safe, while acknowledging in its FAQ that the model can still make mistakes. On the published comparison, a System One task cost $0.000081 and completed in 0.114 seconds, versus $0.013880 and 8.566 seconds for the compared LLM workflow; TypeSafe labels the figures as workflow-specific proof. The site lists Jev at $42 per billion input tokens and describes that price as 238 times lower than Claude Fable 5.1. The product is positioned for automation and machine-native software use, with TypeSafe inviting users to join its waitlist.