Architecture cost review
Jev Router Pricing: Free Router, Variable Bill
"Is Jev Router free?" is the wrong question. The router costs nothing; the models behind it do. What matters is which pool you allow, how often it switches, and what happens when routing fails.
Direct answer
Jev Router charges no fee of its own. OpenRouter bills the model it routes to, at that model's rate, plus the normal credit fee. The saving comes from Jev keeping easy turns on cheap models, adjusting reasoning effort without switching, and avoiding switches that would throw away a cached conversation. Cap the candidate pool with the jev-router plugin, or a hard request can land on a premium model.
Decision summary
| Decision area | What matters |
|---|---|
| Router fee | None — OpenRouter lists typesafe/jev-router pricing as variable; you pay the routed model |
| What Jev decides | Model and reasoning effort for each turn, from difficulty, precision and expected gain |
| Switching rule | Switch only when expected gain beats the cost, including the cache it would lose |
| Pool control | jev-router plugin: models / allowed_models and excluded_models, up to 1,024 patterns each |
| Failure | Jev timeout or invalid output fails the request — no silent fallback to another router |
| Regions | Router models are not available on us.openrouter.ai or eu.openrouter.ai |
What Jev Router charges and what it does not
Set model to typesafe/jev-router and OpenRouter chooses both a model and a reasoning effort for each request. The router's own line is listed as variable because the price is whatever the chosen model costs. The response's model field reports which model served the turn, and that model's rates are what you pay, plus your plan's credit fee.
So the router is free, but the traffic is not. A session that stays on GPT-6 Luna costs Luna rates. A session the router decides needs a heavy model costs Claude Opus 5.5 or GPT-6 Astra rates. The economics depend on how often Jev judges a cheaper model to be enough, and on how you restrict the pool.
Why cache-aware routing is where the money is
Most routers decide turn by turn. Jev Router is built around the observation that switching models discards the provider's prompt cache. The new model must re-read the whole conversation at full input price. OpenRouter says the router keeps a model that works for the rest of the session. It raises or lowers effort instead of switching, and switches only when the expected gain beats the cost, including the lost cache.
The arithmetic explains the design. Re-reading a 200,000-token agent history costs about $0.80 of input on Claude Opus 5.5 at $4 per million, or $0.40 on GPT-6.1 Sol at $2. A naive router that alternates between them on every turn can spend more on re-reads than it saves by sending easy turns to the cheaper model.
Restrict the pool before you trust the bill
By default Jev Router picks from a curated pool and chooses the cheapest candidate that clears the bar for the task. Pass the jev-router plugin with models or allowed_models to narrow that pool, and excluded_models to rule models out. Each list takes up to 1,024 exact slugs, dated revisions, wildcards or latest-family aliases.
The rules have sharp edges. An include list that matches nothing is ignored, and the router falls back to the whole pool. Exclusions are never ignored; if they remove every candidate, the request fails with a 404. Lists can also cap the tier. When a hard request needs a stronger tier than you allowed, the router uses the strongest tier left and reports list_tier_cap in the metadata.
Failure behaviour is a cost decision
If the Jev call times out or returns invalid output, the request fails rather than falling back to another router. That is the opposite of the Auto Router, which degrades to a default model set so a request never fails because routing broke. A strict failure keeps you from paying for a model you did not intend to use. But your client must retry, which adds latency and code.
Budget for that retry path. If you run unattended agents, wrap the call so a routing failure retries once on a fixed model you have priced, then alerts. Count those retries in cost per completed session, or the router will look cheaper than it is.
Worked session cost
Take a 30-turn coding session averaging 40,000 input and 2,000 output tokens per turn. Run entirely on GPT-6.1 Sol with no caching, that is 1.2 million input and 60,000 output tokens: $2.40 plus $0.60, or $3.00. The same session entirely on Claude Opus 5.5 costs $4.80 plus $1.20, or $6.00. Prompt caching lowers both, because cached reads on Sol cost $0.10 per million.
A router earns its keep when it runs most turns at the cheap end and escalates only the one hard debugging turn without abandoning the cache. If it escalates early and stays escalated, the session costs Opus rates. If it switches back and forth, re-reads erase the saving. Enable router metadata and inspect resolved_models per turn before judging the bill.
Region, privacy and procurement notes
Router models, including Jev Router and openrouter/auto, are not available on OpenRouter's regional domains. Teams that must keep inference in the US or EU through us.openrouter.ai or eu.openrouter.ai need a fixed model on those endpoints. They cannot rely on the router.
On privacy, OpenRouter says Jev reads only the conversation text, never attachments, under zero-data-retention terms, and requests sent with zdr set to true work with the router. Document that in vendor reviews. The model Jev routes to has its own data policy, so your provider allowlist and ZDR settings still decide where content goes.
What to log in the first month
Send the X-OpenRouter-Metadata: enabled header and store the jev-router pipeline entry for every request. Its resolved_models field shows the models the router chose, in fallback order. list_fallback shows when your include list matched nothing and was ignored. list_tier_cap shows when your lists capped a hard request at a weaker tier. max_fallback reports when a request that would have had an expert advisor was served by a deep-tier model alone.
Join those fields with usage cost, latency, session ID and whether the task succeeded. After a month you can answer the questions that decide the bill. What share of turns ran on each tier? How many sessions switched models, and what did the re-reads cost? Did capped requests fail more often? That data tells you whether to widen the pool, tighten it or return a workload to a fixed model.
Key takeaways
- →Jev Router charges no fee of its own. OpenRouter bills the model it routes to, at that model's rate, plus the normal credit fee. The saving comes from Jev keeping easy turns on cheap models, adjusting reasoning effort without switching, and avoiding switches that would throw away a cached conversation. Cap the candidate pool with the jev-router plugin, or a hard request can land on a premium model.
- →Start with a restricted pool, enable router metadata, and compare cost per completed session against your current fixed model on the same tasks before moving all traffic.
- →OpenRouter lists the router's own price as variable because you pay the routed model; the 82%-more-tasks figure is OpenRouter's benchmark claim, not an independent test.
How this page was prepared
This September 2026 cluster uses OpenRouter's live models API, OpenRouter and TypeSafe documentation, and the Laya model cards, all checked on 1 October 2026. Vendor benchmark claims are attributed, third-party benchmarks are labelled as such, and every cost example states its token assumptions. We did not run a private benchmark for these pages.
Frequently asked questions
Is Jev Router free?
The router adds no fee of its own. You pay the rates of the model it routes each turn to, plus your OpenRouter credit fee.
Can I stop Jev Router from using expensive models?
Yes. Use the jev-router plugin with models or allowed_models and excluded_models, for example excluding anthropic/claude-opus*. If exclusions remove every candidate, the request fails with a 404 instead of routing elsewhere.
What happens if Jev fails during routing?
The request fails rather than falling back to another router, so your client should retry on a fixed model you have priced.
Does Jev Router work with in-region routing?
No. Router models are not available on us.openrouter.ai or eu.openrouter.ai, so regional traffic needs a fixed model.