TypeSafe AI released Jev this week, a model the company calls its first "System One Model" because it only makes decisions, classifications, routing calls or scores rather than generating text, according to the company's website. Diogo Almeida, a former OpenAI researcher who helped build ChatGPT and its reinforcement-learning-from-human-feedback pipeline, founded TypeSafe AI two years ago with co-founders Erik Gafni and Sasha Sheng.
TypeSafe trains Jev with a technique it calls reinforcement learning for calibrated decisions, producing typed outputs paired with confidence scores instead of free-form prose. The company said Jev ran 193.6 times faster and 444.6 times cheaper than a comparable large language model on structured decision tasks, priced at $42 per billion input tokens.
Newsletter Latent Space reported that within two days of release, developers had already built at least six similar "System One" models of their own. A Claude Code plugin called fast-jev-compaction also appeared on GitHub, using the same approach to score which parts of an agent's context are worth keeping during compaction.
Vercel engineer Pranit Sharma and Bryo AI chief technology officer Nikhil Mudholkar told TechCrunch that Jev matched or beat larger models on accuracy for their classification workloads at a fraction of the cost. TypeSafe's benchmark numbers come from the company's own testing; independent verification is not yet available.
Most of what an agent spends tokens on isn't reasoning, it's deciding: is this input valid, which tool runs next, does this output pass a check. A model priced and trained for exactly that narrow job, instead of a general-purpose LLM doing it as a side effect, changes the cost math for anyone running agents at volume, if the numbers hold up outside the company's own tests.