Compare · Explainer · Last verified 2026-10-28

Jev Router on OpenRouter: what it does, what it costs, how to set it up

Short answer

Jev Router (typesafe/jev-router) is an OpenRouter model slug that uses TypeSafe’s Jev decision model to choose a model and a reasoning effort for each request, then hands the request to that model. It has no fee of its own: you pay the price of whichever model serves the request. It was released on September 25, 2026, works on Chat Completions, Responses and Messages, and is available only in OpenRouter’s global data region. Checked against OpenRouter’s docs and model page on 2026-10-28.

• Use it when you send a mix of easy and hard requests and don’t want to pick one model and one reasoning effort for all of them.
• Skip it when you need EU in-region routing, a fixed model for evaluation or compliance, or a predictable per-request price.

Key facts

Setup: one model string

Send typesafe/jev-router as the model. The response’s model field tells you which model actually answered. This is OpenRouter’s documented minimal request:

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": "Explain the difference between a mutex and a semaphore."}
    ]
  }'

To set a cost tier or model preferences, add the optional jev-router plugin. Every field except id is optional, and unknown fields return 400 (for example, allowed_model is not an alias).

{
  "model": "typesafe/jev-router",
  "plugins": [{
    "id": "jev-router",
    "cost_tier": "high",
    "models": ["anthropic/*", "google/*"],
    "excluded_models": ["anthropic/claude-opus*"]
  }],
  "reasoning": {"effort": "high"},
  "messages": [{"role": "user", "content": "Analyze the tradeoffs in this design."}]
}

Plugin options and the rules that matter

• cost_tier picks the routing policy (low: lower estimated cost; medium: balance; high: more capable models when the benefit justifies the cost). It is a preference, not a price cap, and changing it can change the model mid-conversation.
• models / allowed_models are include preferences with fallback, not allowlists. If they leave no routable model, Jev retries without them. OpenRouter says not to use them to enforce an author restriction.
• excluded_models is a hard rule. If it removes every candidate, the request fails with 404 instead of routing to an excluded model. Use this for compliance.
• Each list takes up to 1,024 patterns (exact slugs, dated revisions, wildcards such as *flash*, or ~author/family-latest aliases), at most 1,024 characters per pattern and 65,536 per list. Matching is case-sensitive.
• Reasoning effort: reasoning.effort is only a hint; the router may choose a different effort or change it across turns. reasoning.max_tokens and reasoning.mode: "pro" return 400.
• In supported agentic workflows the router can keep a model, adjust its effort, or switch models as the task moves on, and when the policy allows it an “expert advisor” model can assist.
• Don’t combine typesafe/jev-router with another router model in one request (400). Errors: 400 for bad fields, regional routing or combined routers; 404 when no eligible model remains or you lack access; 503 when the router is unavailable.

See what the router picked

Send X-OpenRouter-Metadata: enabled and read the jev-router stage of openrouter_metadata.pipeline. It lists the resolved models in fallback order, candidates with effort, capability tier and estimated cost (an estimate, not the billed cost), the policy’s reason, and whether the default or agentic pipeline ran.

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-OpenRouter-Metadata: enabled" \
  -d '{
    "model": "typesafe/jev-router",
    "plugins": [{"id": "jev-router", "cost_tier": "medium"}],
    "messages": [{"role": "user", "content": "Reply with ok"}]
  }' | jq '.openrouter_metadata.pipeline[] | select(.name == "jev-router") | .data | {resolved_models, candidates, reason, pipeline, advisor_model}'

What we ran on 2026-10-28: the public endpoints call needs no key and returned HTTP 200 with the router’s metadata; a chat call without a key returned 401. We did not send paid requests, so we have no first-hand routing results.

curl -s https://openrouter.ai/api/v1/models/typesafe/jev-router/endpoints
# HTTP 200
# {"data":{"id":"typesafe/jev-router","name":"TypeSafe: Jev Router","created":1790363560,
#  "architecture":{"tokenizer":"Router","modality":"text+image+file+audio+video->text", ...},
#  "endpoints":[]}}

Where requests land

OpenRouter’s model page lists the top models used by Jev Router. On 2026-10-28 it showed DeepSeek V4.1 Flash (25.0%), GPT-6.1 Sol (16.1%), GPT-6 Luna (8.6%), GLM 5.3 Flash (8.0%) and GLM 5.3 (7.8%), with the rest spread over other models (34.6%). These are OpenRouter’s aggregate figures across all users, not a promise about your traffic. [OpenRouter data; changes daily]

Jev Router vs OpenRouter Auto Router

Pick Jev Router if you want the router to also set reasoning effort per request and you can live with global-region routing. Pick Auto Router if you want a hard price cap through provider.max_price, more cost bands, or settings saved at the workspace level. If you need repeatable evaluations, call a fixed model instead of either router.

Author routers

When they are in OpenRouter’s catalog, openai/auto, anthropic/auto and other author/auto slugs run Jev with only that author’s models (a hard restriction). The author scope is the model maker, not the host, so an OpenAI model can still be served by another provider such as Azure. openrouter/auto is the separate Auto Router and does not behave this way.

Claude Code and Codex

OpenRouter’s Ori harness starts an agent CLI on OpenRouter and accepts any OpenRouter model ID through --model, and Claude Code can also be pointed at OpenRouter with ANTHROPIC_BASE_URL set to https://openrouter.ai/api. We have not tested Jev Router inside Claude Code or Codex. OpenRouter’s Claude Code guide says Claude Code on OpenRouter is only guaranteed to work with the Anthropic first-party provider, and Jev Router may send turns to other models, so test before relying on it. [not tested by us]

Checklist before you switch

• Log the response model field on every call so you can see cost and quality per routed model.
• Put compliance rules in excluded_models, never in models.
• Need EU in-region routing? Jev Router won’t work; use a fixed model with regional routing.
• Need a decision (yes/no, a label, a score) rather than a written answer? Call a decision API directly; see decision model vs LLM and Jev’s own API ($0.042 per 1M input tokens).
• Guarding agent tool calls instead of choosing models? That is OpenRouter Armor, also built on Jev.

FAQ

Methodology

Every fact on this page comes from OpenRouter’s documentation, model page or public models API, checked on 2026-10-28. Items marked in amber are third-party aggregate data or things we did not test. We made no paid requests. AgentAPIList has no affiliate or sponsor relationship with OpenRouter or TypeSafe. See our methodology.

Sources (checked 2026-10-28)

• OpenRouter docs: Jev Router
• OpenRouter model page: TypeSafe Jev Router
• OpenRouter models API: typesafe/jev-router endpoints
• OpenRouter docs: Auto Router
• OpenRouter docs: Jev hub
• OpenRouter docs: Claude Code integration
• OpenRouter docs: Ori Harness
• TypeSafe docs: Models (Jev)

Related

• Jev (TypeSafe AI): pricing and limits
• OpenRouter Armor: Jev as a tool-call guard
• OpenAI Decisions vs Jev
• Decision model vs LLM
• All comparisons