TypeSafe AI launches System One model Jev for structured automation
Original: Introducing System One Models and Jev
Why This Matters
If the speed and cost claims hold, structured-output models could replace LLM calls in the majority of production automation pipelines.
TypeSafe AI, founded by ex-OpenAI researcher Diogo Almeida, announced Jev on September 15, 2026 — a new class of frontier model built for fast, structured decisions. Jev runs at 70–500ms end-to-end, costs $0.042/MTok for input, and outputs type-safe values with calibrated confidence scores, never generating free-form text.
TypeSafe AI emerged from two years of stealth to release Jev, the company's first 'System One Model' — a category it defines as models optimized for structured, automated decisions rather than human-facing chat. Founder Diogo Almeida previously worked at OpenAI on the instruction-tuning methods behind ChatGPT, and says that experience convinced him chat-style LLMs have a fundamental gap: they're too slow and too unreliable for software integration at scale.
Jev is trained with a new method called Reinforcement Learning for Calibrated Decisions (RLCD), which differs from RLHF and RLVR by targeting epistemically honest probability estimates rather than human preference or verifiable task rewards. The architecture uses a parallel sampler — generating all outputs in a single query rather than token-by-token — which TypeSafe says makes it 40x–200x faster than existing frontier LLMs on comparable tasks.
Pricing is sharply below current market rates: $0.042 per million input tokens, with output tokens described as 'too cheap to meter.' Response times of 70–500ms compare to 3–329 seconds for standard frontier models. Crucially, Jev cannot produce free-form strings — all outputs are typed structured values defined in advance, eliminating hallucinations and parse failures. Every answer ships with a calibrated confidence score. Jev is now available in early access.