TypeSafe AI's Jev Returns Typed Decisions, Not Text
TypeSafe AI shipped a hosted model on September 19, 2026 that answers typed questions with a choice, a score or a probability plus confidence.
In short
Jev is a hosted TypeSafe AI model that takes unstructured state plus a typed question and returns a machine-readable decision with a confidence value instead of prose.
At a glance
- Released September 19, 2026 as a hosted API only, at api.typesafe.ai/v1/systemone; the weights stay closed.
- Three question types: Choice, Score and Noul, a probability between 0 and 1 — all runnable in parallel on one state.
- Pricing is $42 per billion input tokens, with output tokens free. Access runs through a waitlist.
- Vendor's own eval: 193.6x faster than GPT-5.6 Terra (0.114s against 8.566s) and 444.6x cheaper.
- The awesome-jev-tools list catalogs 212 entries across 13 categories, MIT licensed, 219 stars.
Jev does not answer prompts. It answers questions that arrive with a declared answer type: the service takes unstructured state plus a typed question and hands back a choice, a score or a probability, each carrying a confidence value. TypeSafe AI introduced the model on September 19, 2026.
Three question types, no prose
Choice picks from a supplied list. Score places the state on ordered levels. Noul returns a truth judgment as a probability between 0 and 1. The vendor says all three can run in parallel against the same state at little latency cost, and quotes 70 to 500 milliseconds end to end — the band that matters for routing, classification and pre-flight checks before a tool call.
Billed on input alone
The rate is $42 per billion input tokens, output tokens free — a sensible shape for a model whose answer is a handful of fields. The endpoint is api.typesafe.ai/v1/systemone, with Python and JavaScript SDKs. Access sits behind a waitlist. No weights are distributed, and self-hosting is not offered.
The benchmarks belong to the vendor
TypeSafe AI's own workflow evals put Jev 193.6x faster than GPT-5.6 Terra (0.114 seconds against 8.566) and 444.6x cheaper. The company wrote those workflows itself and explicitly frames the results as the high end of real use; it also concedes it cannot prove the pricing is unsubsidized. Vercel CEO Guillermo Rauch cites 18x faster p95 than GPT Luna. No independent measurements have been published so far.
A catalog running ahead of access
The MIT-licensed awesome-jev-tools list already carries 212 entries across 13 categories and sits at 219 stars. Infrastructure and SDKs lead with 33 entries, agent decisions follow with 29, and classification and guardrails hold 20 each. The list gives no release date for Jev itself; that date comes from press coverage.
What stays unverified
Architecture and parameter count are undisclosed. The widely repeated “zero hallucinations” line covers schema conformance only — an answer that matches the schema can still be factually wrong. Teams that put Jev in front of consequential actions are relocating the risk rather than removing it.
FAQ
What is TypeSafe AI's Jev?
A hosted model that answers a typed question about unstructured state, returning a choice, a score or a probability between 0 and 1, each with a confidence value.
How much does Jev cost?
$42 per billion input tokens, with output tokens not billed. Access is currently limited to a waitlist.
Are Jev's weights available for self-hosting?
No. The model is proprietary and reachable only through api.typesafe.ai/v1/systemone; architecture and parameter count have not been published.
