Introducing System One Models & Jev — TypeSafe AI Blog

Claims

1. eval-caveats
   TypeSafe discloses that its workflow evaluations were written by its own capabilities team; that reference answers were the average of GPT-6 Astra and Fable 5.1, which may favour OpenAI and Anthropic models; that LLMs were run through TypeSafe's own wrapper constraining them to structured decisions; that latency was largely measured from West Coast laptops; and that it cannot prove its pricing is not subsidised.
2. function-call
   TypeSafe describes Jev as a function call backed by frontier-level intelligence: unstructured state goes in, typed probabilistic decisions come out.

   "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out"

   Quote language: en
3. hallucination
   TypeSafe claims Jev cannot hallucinate and never makes type errors; its reported 0% hallucination rate is described as a guarantee that follows from matching the schema, not an empirical result.

   "The model never makes type errors."

   Quote language: en
4. naming
   "System One" draws on Daniel Kahneman's distinction between System 1 and System 2 thinking; "Jev" is named after William Stanley Jevons, reflecting TypeSafe's expectation that cheaper intelligence will raise demand for it, as more efficient steam engines raised demand for coal.
5. speed-cost
   TypeSafe states an end-to-end response time of 70–500 ms, 40 to 200 times faster than frontier LLMs on System One-style queries; input priced at $0.042 per million tokens with free output; and homepage headline figures of 193.6 times faster and 444.6 times cheaper, derived from its own workflow evaluations and expected to be at the higher end of real-world gains.

   "End-to-end response time is 70ms-500ms for TypeSafe."

   Quote language: en
6. stack
   TypeSafe says it built a new model architecture, a parallel sampler and a training method called Reinforcement Learning for Calibrated Decisions (RLCD), and that Jev generates all outputs in a single query rather than token by token. The post gives no technical detail on the architecture or RLCD.
7. uses
   TypeSafe names intended uses as AI-powered workflows (classify, route, score, extract, branch), map-reducing over large datasets, real-time applications, and verification tasks such as judging, guardrailing and detecting jailbreaks in the prompts and outputs of LLMs. Jev was released in early access with a waitlist.