September 19, 2026
A dedicated model arrives for high-volume classifications: Jev costs $0.042 per 1 million tokens
On September 15, 2026, TypeSafe AI released Jev, a System One model for fast structured decisions. Version jev-1.13.0 takes state and questions, then returns Choice, Score, or Noul with probabilities and confidence scores. TypeSafe claims 70–500 ms latency versus 3–329 seconds for the highest-tier models.
Classification and rule-based selection were assigned to large language models. Jev takes state and questions with specified types, processes questions in parallel, and returns a structured response rather than text or code.
In practice. In an experiment described by Flavio Copes, Jev classified 1018 AI articles across 24 topics for $0,08 at a median latency of 256 ms. Another experiment classified 98 thousand listings in 10 minutes.
Diogo Almeida writes that he contributed at OpenAI to methods that became the foundation of ChatGPT. For Python, install the package with `uv add typesafe-sdk`, create a key in the TypeSafe console, set `TYPESAFE_API_KEY`, and call `TypeSafeClient.system_one`.
The limit for a single request is 64k tokens, and state can take up 32k tokens plus the longest question. As of September 16, Jev was available only through a waitlist, with no independent replication of the benchmarks.
The next benchmark will be independent measurements of Jev, which did not yet exist as of September 16.
