Introducing System One Models & Jev (TypeSafe AI, 2026-09-15)

Launch post by Diogo Almeida, founder of typesafe-ai, dated 15 September 2026. Arrived by Telegram 2026-09-16. It announces a model class ("System One models"), a first model (Jev), a training method it calls RLCD (Reinforcement Learning for Calibrated Decisions), early access behind a waitlist, and a price. Companion pages read: the workflow evals at evals.typesafe.ai and the listing of a system-one-adapter-python repository (not opened).

What is being sold

Not a chat model. Jev is pitched as "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." The caller declares the output type in advance (yes/no, a choice with cardinality up to 255, a score), and the model returns that value with a probability attached. Three consequences the post draws: the model "never makes type errors" (the schema is enforced, so the type-error rate is 0% by construction), it "can't hallucinate" (it cannot emit free text), and it "always communicates confidence." The architecture claim is a parallel sampler — outputs are generated at once rather than token by token — which is where the speed claim comes from.

The name is daniel-kahneman's System 1 (fast, automatic) from Thinking, Fast and Slow, filed in psychology-wiki under heuristics-and-biases; Jev is William Stanley Jevons, whose paradox — a cheaper engine burns more coal — the post invokes as the intended effect of a cheaper decision.

The numbers, as the vendor states them

End-to-end response 70 ms–500 ms against 3 to 329 s for frontier LLMs on the same workflows; "40x–200x faster for the same levels of frontier intelligence."
Input $0.042 per million tokens ("$42 per billion"); output free ("too cheap to meter"). LLM comparators quoted at $0.20–$10 per million input, sourced from OpenRouter.
On the workflow evals, headline ratios of 193.6× faster, 444.6× cheaper, which the post itself says are "on the higher end of real world gains."

What the evals actually are

evals.typesafe.ai runs four workflows — security incidents, agent-trace observability, invoice processing, customer service — each decomposed into yes/no ("Noul"), choice and score questions with programmatic decision logic. Metrics: accuracy, cost per case, latency per case. Jev's aggregate line is 67.8% accuracy, $0.0004 per case, 0.4 s. Comparators include Haiku 4.5, Sonnet 5, Opus 5, DeepSeek v4 Flash/Pro and the GPT-5.6 Luna/Sol/Terra tiers, at provider defaults.

Three things bound the result. The reference labels are an average of GPT-6 Astra and Fable 5.1 "at high thinking," with "the harness is correct" stated as an assumption — so accuracy is agreement with two other vendors' models, not with ground truth, and the reference models cannot lose. The workflows were written by TypeSafe's own capabilities team, which the post discloses. And the four workflows are the product's home turf: closed-set classification over structured inputs, which is the case a typed decision model is built for and an LLM is not. That Jev owns the cost/latency Pareto frontier on it is expected; that it reaches 67.8% agreement is the number worth watching, and nobody has run it independently.

Where it sits in this spoke

The nearest neighbour is inception-labs's Mercury (inception-platform-docs): a second provider whose whole pitch is not autoregressive and whose competitive axis is latency. Mercury publishes no benchmark; TypeSafe publishes one it built and graded itself. On price, $0.042 input with free output sits below Mercury's promotional $0.04/$0.15 floor in llm-api-pricing, but the two are not comparable — Jev's output is a typed value, so "output tokens" barely exist.

Tier T3 — vendor launch post; every figure is the vendor's own, on the vendor's own eval, graded against models the vendor chose. The disclosures are unusually complete for the genre, which is why it is T3 and not T4.

Open

Independent accuracy. Growth edge 1 (an independent re-run of any vendor claim) applies in full; the eval harness is public, so a re-run with human labels is possible.
~~Availability and price after early access.~~ Answered 2026-09-19 by jev-ai-sdk-guide: Jev is generally available through Vercel's AI Gateway as typesafe-ai/jev, and the $0.042/1M input with free output held through the transition. The free-output price survived one launch, which is one more than most do.
What "calibrated" means in practice — advanced, not closed, by the same source. Vercel states it plainly: calibration is agreement between predicted probabilities and observed outcomes across many examples, and "doesn't guarantee an individual answer is correct." So the promise is about a distribution, not an answer. Still no calibration curve and no Brier score from anyone.

typesafe-ai · jev-ai-sdk-guide · inception-platform-docs · inception-labs · llm-api-pricing · llm-benchmarks · astra · claude-fable-5-1 · heuristics-and-biases · daniel-kahneman