Jev on OpenRouter: Structured Classification & the Decisions API

Which Jev models are on OpenRouter?

Checked September 29, 2026 against openrouter.ai/typesafe and the public endpoints API. The default GET /api/v1/models catalog included Jev Router and did not include Jev 1.13. The decision model is on its own model page and endpoints URL.

TypeSafe's own models page lists jev-1.13.0 at the same input price, with a larger direct-API window: 64,000 tokens per request, and 32,000 for the state plus the longest question. Gateway context and direct context are different numbers. Quote the one for the path you call.

Decisions API vs chat completions

A decision call sends state plus named questions. Each question is a Noul (yes or no probability), a Choice (one of your options, with a probability for each), or a Score (a position on your rubric). Jev 1.13 does not write the reply. Your code reads the probabilities and applies a cutoff.

OpenRouter documents two surfaces, both billed to the same OpenRouter key:

• Decisions API: POST https://openrouter.ai/api/alpha/decisions. Use this from HTTP or the OpenRouter SDKs. The Jev tutorial is the official walkthrough.
• System One API: POST https://openrouter.ai/api/v1/systemone. Point an existing TypeSafe SDK at https://openrouter.ai/api. Model id jev-1.13 is routed as typesafe/jev-1.13. See OpenRouter's TypeSafe SDK note.

The Jev 1.13 model page says chat-completions SDKs will not work with it. Chat completions are the shape for Jev Router, not for this model.

Shortest Decisions shape, with the key left in the environment:

curl https://openrouter.ai/api/alpha/decisions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "typesafe/jev-1.13",
    "state": "Checkout shows a blank page after Pay.",
    "questions": {
      "is_bug": {
        "type": "noul",
        "instructions": "Is the customer reporting a software defect?"
      }
    }
  }'

Question criteria, response fields, and SDK samples stay in the OpenRouter Jev docs and TypeSafe's System One concept. The lab export at /docs shows one pinned gate. It is not a general Decisions client.

The 27% classification claim

OpenRouter stated on September 26, 2026 that Jev took 27% of weekly request volume in the classification category, nearly twice the share of DeepSeek V4 Flash, which had held the top spot. Treat that as a platform claim about classification requests. It is not a JevLab measurement, and the post does not name Jev Router.

Source: OpenRouter, September 26, 2026. The category is classification request volume on OpenRouter. It is not token share across every model, and it is not a measurement of typesafe/jev-router.

OpenRouter also published its own Banking77 comparison on September 22, 2026: Jev 1.13 at 81.0% accuracy and Claude Opus 5 at 84.4% on 3,080 utterances, with Jev faster and cheaper in that run. That writeup is OpenRouter's measurement. JevLab's lab is a five-label routing cutoff on a candidate fixture. It is not that dataset, and no reviewed accuracy is claimed here.

Production tips for a classification gate

• Pin jev-1.13.0 on TypeSafe, or typesafe/jev-1.13 on OpenRouter. Aliases move, and a cutoff is calibrated to one score distribution.
• Read Choice and Score confidence as a gate, not as accuracy. The three-band threshold guide separates auto-adopt, a second look, and handoff.
• A schema-valid label can still be the wrong label. Schema determinism is not factual truth.
• On timeout, HTTP 429, or HTTP 529, fail closed. Key handling and the documented limits are on Jev API access.

Sources checked on September 29, 2026

• OpenRouter: TypeSafe models
• OpenRouter: Jev 1.13
• OpenRouter Jev documentation
• TypeSafe models
• OpenRouter classification-volume post

Frequently Asked Questions