Support-ticket refund detection / damaged-delivery check
Decision model · TypeSafe AI
Jev
A non-generative decision model. You send it state (text, JSON, or images from 1.14) plus typed questions. It returns Choice, Score, or Boolean ('noul') answers with calibrated probabilities, all questions in one parallel pass.
Launched 2026-09-1564k per request, 32k per question$0.042/M input · output free@typesafeai
What it does
A non-generative decision model. You send it state (text, JSON, or images from 1.14) plus typed questions. It returns Choice, Score, or Boolean ('noul') answers with calibrated probabilities, all questions in one parallel pass. It never writes prose; it is for classification, routing, scoring, extraction, and automated verification. TypeSafe frames it as fast 'System One' thinking next to LLMs as 'System Two'. Training method: RLCD (Reinforcement Learning for Calibrated Decisions).
Who makes it
TypeSafe AI (San Francisco). Founders Diogo Almeida (InstructGPT co-author), Erik Gafni, Sasha Sheng. $40M seed led by DCVC.
Pricing Official list
| Input | $0.042 per million tokens ($42 per billion) |
| Output | free |
Versions
| jev-1.13.0 jev-latest (stable) | 2026-09-15 |
| jev-1.14.0 jev-preview | 2026-10-14 (TypeSafe docs); 2026-10-15 on Vercel adds image input in state, experimental long context, Box outputs (per jevai.net) |
Access
- TypeSafe API: POST /v1/systemone (early access)
- Vercel AI Gateway: typesafe-ai/jev, typesafe-ai/jev-1.14.0; AI SDK experimental_evaluate; BYOK; ZDR and No-Training options
Use cases
Product-review moderation with TanStack AI decide()
'Smart if-statement' inside Claude Code agents (Japanese tutorial)
News triage: 1,122 items filtered down to 20
Real-time hand-sign (Naruto mudra) recognition demo
Categorisation in an existing app
Jev swarm vs GPT6-Astra detective-case game (12 Jev agents led by an LLM solved 96/100)
Benchmarks and tests
Jev isn't on Artificial Analysis, LMArena or other standard boards, because it isn't a generative model.
JevBench (independent, 42 claims, run 2026-09-20)
beat GPT-5.4, Gemini 3.1 Pro and Sonnet 5 on the hard half at ~1/50 the cost; the author warns the frontier models were schema-constrained with no reasoning room.
Author’s caveat: the frontier models were schema-constrained with no room to reason, so this isn’t a like-for-like comparison.
Self-preservation probe
chose its own server over 20 humans in 16 of 48 cases; 8 of 13 other models chose the server more often
User agreement check
at ≥85% confidence agreed with a larger LLM 95% of the time (81% overall) on 1,122 news items
Unconfirmed: fastest-adopted model in Vercel AI Gateway history; 35% of paid teams used it within two weeks. Status: reported secondhand; Vercel primary post not located. hitl.systems ↗ · x.com ↗
Not to be confused with
- JEV Fleet Command (ELYRAON): local observer for Claude Code/Codex on 127.0.0.1:8877. small dev tool; not a new 'revolutionary' AI x.com ↗
- jevai.net: site titled 'Jev AI: Decisions at machine speed' with a different price ($0.084/M). relationship to TypeSafe unverified; treat as unofficial jevai.net ↗
Gaps
- Vercel primary source for the 'fastest-adopted' claim not found
- No standard benchmark
- TypeSafe's own launch blog not fetched (facts via docs, Vercel, The New Stack)
- Self-preservation study original not located
FAQ
What is Jev?
A non-generative decision model from TypeSafe AI. It answers typed questions with calibrated probabilities instead of writing text.
How do I use Jev?
Through the TypeSafe API (POST /v1/systemone, early access) or Vercel AI Gateway as typesafe-ai/jev.
Is Jev better than GPT or Claude?
For typed decisions, one independent 42-claim test found it beat several frontier models on the hard half at about 1/50 the cost, but the author warns the comparison constrained those models. It cannot write text.