Jev API Fast, type-safe structured decisions
The Jev API is the developer interface to Jev, the first System One model from TypeSafe AI, released on September 15, 2026. Instead of generating text, the Jev API reads your application state and returns type-safe answers — a choice, a score, or a probability — each with calibrated confidence, in 70 to 500 milliseconds.
model: jev-latest · status: early access
What is the Jev API?
Jev turns fuzzy judgment into typed, testable function calls.
A request carries a state —
plain text or structured JSON — and a map of questions, each built
from one of three primitives. The response carries an answer for
every question, its probability distribution, and a confidence value,
so your code — not a prompt — owns the control flow.
Because Jev was trained for decisions rather than text generation, there is no hidden chain of thought and no free-form output to parse or defend against. Answers cannot violate your declared schema, the same state is evaluated for all questions in parallel, and every result arrives with a confidence number that is meant to be used as a threshold. What that costs you — the generation it cannot do — is what Jev API vs GPT is about.
Type-safe by construction
No malformed JSON, no undeclared options, no prompt wrangling for structure.
Calibrated probabilities
Confidence is an explicit training objective, so thresholds behave like they should.
Parallel by design
Every question on a state is evaluated in one call — fan out freely.
Three question primitives
Every question you send is one of three typed shapes. Compose them freely in a single request — the response maps each answer back by key.
Noul
Boolean probability
Ask a yes/no question of the state and get the probability that the answer is true — the smallest, fastest building block for filters, gates, and guardrails.
{ noul: number } // 0–1 Choice
One of a fixed option set
Declare the criteria for every option; the API picks one and returns the full probability distribution plus confidence — perfect for routing and tool selection.
{ choice, probabilities, confidence } Score
Ordered rubric grading
Grade the state against an ordered scale you define, with a per-level distribution and a legend your code can consume — triage, risk, and quality scoring.
{ score, probabilities, confidence, legend } Your first Jev API request
One HTTPS call classifies the state on two questions at once — a Noul gate and a Choice route, in a single round trip.
$ curl -X POST https://api.typesafe.ai/v1/systemone \
-H "Authorization: Bearer $TYPESAFE_API_KEY" \
-H "Content-Type: application/json" \
-d '{\
"model": "jev-latest",
"state": "Help! My payouts have been failing for 3 days.",
"questions": {
"is_urgent": {
"type": "noul",
"instructions": "Does this convey urgency?"
},
"department": {
"type": "choice",
"instructions": "Which team should handle this?",
"criteria": {
"billing": "Payments, invoicing, refunds",
"technical": "Bugs, outages, integrations",
"sales": "Pricing, upgrades, new accounts"
}
}
}
}' # → 200 OK, every answer in parallel
# abridged response
{
"is_urgent": { "noul": 0.97, "confidence": 0.93 },
"department": {
"choice": "technical",
"probabilities": {
"billing": 0.04, "technical": 0.83, "sales": 0.13
},
"confidence": 0.88
}
}- 1
Join the early-access waitlist
Access is rolling out in waves. Request it at typesafe.ai — most developers are admitted within a day or two. typesafe.ai
- 2
Create an API key
Sign in to the console and create a key. Keys authenticate with a standard bearer token. console.typesafe.ai
- 3
Install an official SDK
Python 3.10+: pip install typesafe-sdk · Node 20+: npm install @typesafe-ai/sdk. Both read TYPESAFE_API_KEY from the environment and default to jev-latest.
- 4
Send your first question set
Post a state and a map of questions to the systemone endpoint. Every question is evaluated in parallel on the same state.
What teams build with the Jev API
The pattern is always the same: fixed options, high frequency, and a decision that code — not prose — has to act on. Most of these replace a call you are already making to a language model or a trained classifier, and the comparisons work through which.
Routing & triage
Send support tickets, prompts, or jobs to the right team, model, or agent — chosen from a fixed option set with confidence attached.
Guardrails & safety
Gate every prompt and output: injection detection, policy checks, and escalation whenever confidence drops below your threshold.
Agent supervision
Let a Jev call pick the tool, compact the context, or decide when an agent is stuck — while the LLM handles generation and explanation.
Real-time decisions
At 70–500 ms a Jev call fits inside control loops: game agents, browser automation, voice, and robotics.
Search & reranking
Score candidates against a rubric to rerank BM25 shortlists before they reach a downstream LLM.
Batch classification
Tag, score, and filter millions of items in parallel. Output tokens are free, so volume stops being a cost problem.
The Jev API at a glance
- Endpoint
- POST https://api.typesafe.ai/v1/systemone
- Authentication
- Authorization: Bearer <API_KEY>
- Models
- jev-latest → jev-1.13.0 · jev-preview
- Context
- ~64k tokens for state + all questions (~32k for state + longest question)
- Latency
- 70–500 ms typical; parallel questions add almost no latency
- Rate limits
- ~250k tokens/s and 1200 requests/min, adjusted during early access
- Pricing
- $0.042 per million input tokens · output tokens free
- Error codes
- 401 auth · 422 validation · 429 rate limit · 529 overloaded
Figures reflect the early-access release and are adjusted over time — always confirm against the official models page . The pricing and latency figures are examined against a generative model in Jev API vs GPT.
Jev API questions, answered
What is the Jev API?
It is the REST interface to Jev, the first System One model released by TypeSafe AI on September 15, 2026. You send a state plus a set of typed questions; the API returns type-safe answers — a boolean probability, a choice, or a rubric score — each with calibrated confidence, in 70–500 ms. It does not generate text.
Can the Jev API hallucinate?
Structurally, no: every answer is constrained to your declared schema, so the API cannot return a malformed object, an undeclared option, or a type error. But a model can still assign high confidence to a wrong-but-valid option. Treat outputs as probabilities — set confidence thresholds and keep a fallback path for low-confidence results.
How much does the Jev API cost?
Input costs $0.042 per million tokens and output tokens are free, which makes high-volume decision workloads — routing, classification, scoring — dramatically cheaper than generative LLM calls.
How fast is the Jev API?
Official latency is 70–500 ms per request, and community measurements typically land between 70 and 300 ms. All questions on one state are evaluated in parallel, so adding questions to a request adds almost no latency.
How do I get access to the Jev API?
Jev is in early access. Join the waitlist at typesafe.ai, then create an API key in the console at console.typesafe.ai once admitted. Gateway model IDs are also available: typesafe-ai/jev on Vercel AI Gateway, typesafe/jev on Cloudflare Workers AI, and the official SDK on Netlify AI Gateway.
Which SDKs does the Jev API support?
Official SDKs are typesafe-sdk for Python 3.10+ and @typesafe-ai/sdk for JavaScript and TypeScript on Node 20+. Community clients exist for Rust, Ruby, Go, and Elixir, and an agent skill lets coding assistants call the API directly.
How the Jev API compares
Jev is not an LLM, so most of these are not like-for-like. Each page says what the other thing does well, what Jev replaces, and where Jev is the wrong choice.
- Jev API vs Claude Sonnet 5 Claude reasons about a decision and can explain it. Jev makes the decision and cannot. Most agent stacks need Claude doing the first job more than they need it doing the second.
- Jev API vs GPT-5.6 Terra Not rival products — Jev replaces the classification call, GPT keeps the reasoning call. The stack usually wants both.
- Jev API vs Instructor and Outlines Instructor and Outlines constrain a model that is still generating. Jev is a model that was never generating. Pick by whether you need the generation at all.
- Jev API vs RouteLLM and Not Diamond RouteLLM and Not Diamond are routing infrastructure with a learned selector inside. Jev is a selector. The interesting question is which belongs in the loop.
- Jev API vs Traditional ML classifiers A trained classifier is cheaper per call, explainable and yours. Jev is faster to start, survives label changes, and needs no data. The deciding question is whether your labels are stable.
Official Jev API resources
This page is a reference — the source of truth is TypeSafe AI's own documentation.
- Documentation center
- API reference
- Quickstart
- Models & pricing
- Playground
- Console & API keys
- Launch announcement
- Workflow evals
- llms.txt index
About this page
JevAPI.dev is an independent developer reference written and maintained by engineers who build on Jev. It is not affiliated with or endorsed by TypeSafe AI; Jev and TypeSafe are trademarks of TypeSafe AI. Facts on this page were last verified against the official documentation on .
Put the Jev API in front of your control flow
Join the early-access wave, then bring routing, guardrails, and scoring in-house — one typed question at a time.
Still weighing it against what you run today? Compare the Jev API with GPT, Claude, routing tools and trained classifiers.