Jev API Guide: Endpoint, Model Names, SDKs and Pricing

Jev is served over a single HTTP route, and every official integration — the Python and JavaScript SDKs, Pydantic AI, LiteLLM, LangChain, Vercel's AI SDK, the Netlify AI Gateway — is a wrapper around it.

The endpoint

POST https://api.typesafe.ai/v1/systemone
Authorization: Bearer $TYPESAFE_API_KEY
Content-Type: application/json

{
  "model": "jev-latest",
  "state": "Customer emailed twice this week about a failed refund.",
  "questions": {
    "team": {
      "type": "choice",
      "instructions": "Which team should handle this?",
      "criteria": {
        "billing": "Invoices, payments and refunds",
        "technical": "Bugs, outages and integrations",
        "other": "Anything outside those categories"
      }
    }
  }
}

The response echoes the model version that actually answered, the chosen value, the full probability distribution, and a confidence number you can threshold on:

{
  "model": "jev-1.13.0",
  "answers": {
    "team": {
      "type": "choice",
      "choice": "billing",
      "probabilities": {"billing": 0.85, "technical": 0.08, "other": 0.07},
      "confidence": 0.82
    }
  },
  "usage": {"input_tokens": 312, "output_tokens": 48}
}

Model names

jev-latest — floating alias, the one to use while experimenting. As of September 2026 it resolves to jev-1.13.0.
jev-1.13.0 — pin this the moment you tune a threshold. A floating alias can silently move your answers.

SDKs and integrations

Path	Package / model string
Python SDK	Official TypeSafe client, reads `TYPESAFE_API_KEY`
JavaScript SDK	`@typesafe-ai/sdk`, helpers like `choice()`
Pydantic AI	`Agent('typesafe:jev-latest')` with a typed output model
LiteLLM	Pass-through route `/typesafe/v1/systemone`
Vercel AI SDK	`typesafe-ai/jev` via `experimental_evaluate`; confidence in provider metadata
Netlify AI Gateway	Same SDK, key injected by the platform
Routers	`typesafe/jev-latest` through gateways such as Requesty
Agent CLIs	`typesafe-ai/skills` plugin marketplace for Claude Code and Codex

Pricing and throughput

Reported pricing is ~$0.042 per million input tokens with output tokens unmetered, because a decision produces almost no output. Developers publishing their own bills describe thousands of calls for a couple of dollars: one widely shared example logged about 5,000 requests — classification, model routing, intent and steering combined — for roughly $2. A browser automation demo was quoted at 7 seconds and $0.0039 for a full flight-search run.

Those are community numbers from the launch week, not a rate card. Check your dashboard before promising anyone a budget, and remember rate limits live on the API key, not on this page.

Practical rules that save you a rewrite

One call, all questions. Asking five things in one request is cheaper and more consistent than five requests.
Do deterministic work in code first. Counts, sums, comparisons, timestamps — compute them, then pass them in as fields.
Give every choice an escape option. If there is no "other" or "unclear" bucket, the model is forced to guess.
Write the criteria as definitions, not as adjectives. Say what belongs in each bucket and what does not.
Pin the version, then log the version that answered. Tuner's rule: the threshold and the model version are one artefact.

Next: working code examples · use cases · running the same pattern offline with Laya

Last updated: 2026-09-21 · sources & corrections