The AI industry spent two years teaching models to chat. Now, TypeSafe AI argues the real breakthrough is teaching them to decide. Their new model, Jev, skips natural language generation entirely. Instead of writing paragraphs, it outputs typed decisions with confidence scores that code can immediately act on. This shift targets the bottleneck in autonomous agents: the need for a human to verify vague text outputs. By forcing structured answers, Jev aims to remove the loop.
Jev operates on what TypeSafe calls a “System One” architecture, borrowing from Daniel Kahneman’s concept of fast, intuitive thinking. Unlike large language models tuned for human preference via RLHF, Jev uses Reinforcement Learning for Calibrated Decisions (RLCD). The API accepts a state and specific questions, returning results in milliseconds. In internal benchmarks, TypeSafe claims Jev is 193.6 times faster and 444.6 times cheaper than GPT-5.6 Terra. A typical request cost $0.000081 and completed in 0.114 seconds, compared to $0.013880 and 8.566 seconds for the competitor. Input tokens are priced at $42 per billion, while output tokens are free.
Early adopters are testing these claims in production-like environments. Vercel CEO Guillermo Rauch noted Jev was up to 18 times faster at p95 latency than GPT Luna for command safety checks. Other developers have built tools for email triage, browser automation, and even StarCraft gameplay. However, the benchmark data comes with caveats. TypeSafe admits the comparisons rely on vendor-run evaluations and averaged reference answers from unreleased models like GPT-6 Astra. They also acknowledge they cannot prove the current pricing is unsubsidized.
The promise of “zero hallucinations” refers strictly to schema matching, not factual accuracy. Answers can still be wrong, but they will always fit the expected data type. This distinction matters for engineers building guardrails. While the speed and cost advantages look compelling on paper, the reliance on proprietary, closed-source architecture means developers must trust TypeSafe’s black box. The real test won’t be in demo videos, but in whether enterprises can tolerate the risk of unverified weights in critical decision pipelines.
