Jev Router, measured: which model it picks and what it costs

What Jev Router is

OpenRouter's listing describes Jev Router as picking the best model and reasoning effort for each request, balancing quality, speed and cost, and adapting as a conversation evolves. It runs on Jev, the System One model TypeSafe launched on September 15, 2026. Jev itself never writes text; in Jev Router it makes one decision — which model — and the model it chose writes the reply.

It accepts the usual chat parameters, including tools, structured outputs and a reasoning setting, and OpenRouter lists text, image, file, audio and video as inputs. TypeSafe has not published which models are in its pool or how it scores them, which is why the rest of this page is a measurement rather than a summary.

How to call Jev Router

Jev Router is reached like any OpenRouter model: the OpenAI-compatible endpoint with your OpenRouter key and the model id. Ask for usage to see what each call cost; the response's model field names the model that actually answered.

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "typesafe/jev-router",
    "messages": [{"role": "user", "content": "Prove that the square root of 2 is irrational."}],
    "max_tokens": 2000,
    "usage": {"include": true}
  }'

Leave max_tokens generous. Jev Router can choose a reasoning model, and reasoning tokens count against the same limit — the one miss in our run below is exactly that.

What Jev Router picked, request by request

The pattern in what Jev Router picked is plain. For short facts, arithmetic, classification, code and even the question in Chinese, Jev Router picked GPT-6 Luna, the cheapest model it used. The French greeting, the translation, the SQL, the table and the migration plan went to DeepSeek V4.1 Flash. Jev Router picked Gemini 3.8 Flash twice, on arithmetic and counting, and once, for the proof, Claude Sonnet 5.5 — the most expensive call in the run, at about a quarter of a cent.

What Jev Router cost

For all 24 requests together Jev Router billed $0.0056, and every call matched the list price of the model that answered it to the last digit, so the router added no fee of its own. Priced on a single model, the same token counts would have cost $0.0468 on Claude Opus 5.5 and $0.0234 on GPT-6.1 Sol — 8.3 and 4.1 times as much. On GPT-6 Luna alone they would have cost $0.0012, a fifth of what Jev Router billed.

That last comparison is the honest one. Jev Router saves money against sending everything to a flagship, not against sending everything to the cheapest model; what it buys you is the escalation — the proof went to a stronger model without anyone writing a rule for it. A different model would also have used a different number of tokens, so read the single-model figures as a guide. To price your own traffic on any model, use the LLM cost calculator.

Where Jev Router went wrong

Once. For "Find the sum of all integers n with 1 ≤ n ≤ 100 that are divisible by 3 or 5. Give the number." Jev Router picked Gemini 3.8 Flash, which spent 385 of its 396 output tokens reasoning and reached our 400-token cap before writing the answer. That is our limit, not a wrong answer — but it is the failure you will see in production if you cap replies tightly, because you do not choose the model and some of the models Jev Router picks think before they answer. Every other request with a checkable answer came back right: 15 of 16.

Jev Router or Jev itself?

Use Jev Router when you want an answer and would rather not maintain routing rules. Call Jev directly when you want the decision: your own list of models or tools as options, your own confidence threshold, and the choice logged in your code, at $0.042 per million input tokens for the decision alone. The model routing use case shows that pattern with measured output, and Jev in an agent applies it to tool calls.

Skip both when every request needs the same model for consistency or compliance, or when your traffic is all easy — in our run, one cheap model would have been cheaper still.

Jev Router questions

Method: 24 prompts, one run each, through OpenRouter on 2026-10-02, with max_tokens at 400. Answers with a fixed result were checked by string match; open-ended ones were not scored, so what Jev Router picked for them is recorded but not graded.