Jev: TypeSafe's System One Model Explained

Not from this page. The fetch returned an HTTP 403 and captured no text, so the source does not document anything directly; this comes from the entry's own notes on it. Those notes record that the eval workflows and the reference answers were written by TypeSafe, which makes parity with GPT-5.6 Terra promising rather than settled.

Mostly typing. The notes attribute Jev's 0% to construction rather than performance, against 0.58% for luna and terra, 5.73% for Opus 5 and 45.5% for Claude Haiku 4.5 in the same vendor table. Tool-call errors reach 17.0% for GPT-5.6 Sol. The page itself did not load, so it does not document how those rates were counted.

The entry describes waitlisted early access rather than open signup, and the documented call is POST https://api.typesafe.ai/v1/systemone with model jev-latest, with Python and JavaScript SDKs listed. The page returned an access error when fetched, so the source does not document waitlist mechanics, queue time or approval criteria.

Run your own gate on your own labels at your own volume. The notes put Jev at $0.0004 per case and 0.4s against GPT-5.6 Terra's $0.0304 and 10.1s on the same four workflows, on vendor-written reference answers. Treat the ratio as a reason to test, not a number to plan capacity around.

More than a stricter API. Jev uses a model architecture with a parallel sampler that produces all outputs in a single query rather than autoregressively, and a training method TypeSafe calls RLCD, Reinforcement Learning for Calibrated Decisions, which optimizes for calibrated probabilities on structured decisions rather than for text humans prefer. That is why calibration is a property of the model rather than something bolted on by prompting.

Jev returns probabilities and no natural-language rationale, and reviewers flag that as the real constraint for debugging and for audits in regulated domains. A bare confidence number will not satisfy a compliance team that needs to know why a loan application scored as it did. The split the article recommends is to keep Jev for the high-volume routing layer and escalate flagged or low-confidence cases to a model that can write the explanation.

Jev bills $0.042 per million input tokens and does not meter output, roughly 1/48th of GPT-5.6 Terra's $2.00 per million input, and Terra adds $12.00 per million output on top. TypeSafe is candid that it cannot prove the price is not subsidized and expects prices to fall rather than rise, so the durability of the gap is the open question rather than the current number.