header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP

OpenRouter integrates Jev into smart routing: calculate cache loss before switching models.

Beating AI News Flash: OpenRouter has launched Jev Router, integrating TypeSafe's Jev into model routing. Before each request is sent, Jev assesses task difficulty, precision requirements, and whether the current model and reasoning tier are sufficient, then decides whether to continue with the same setup, increase reasoning effort, or switch to a different model.


OpenRouter's original Auto Router was more focused on selecting models by task type. It first determines what kind of task it is, then references how various models have been used for similar requests over the past 7 days to choose a route. Jev Router goes further by evaluating how difficult the current round actually is. Simple tasks can be downgraded, while hard problems can have reasoning effort increased; if the same model can still handle it, there's no rush to switch.


There's also a very practical issue here: context caching. In long conversations, once the model is switched, the already-cached context typically cannot be directly reused, and the new model may need to reprocess the entire chat history. Jev Router factors this loss into its calculation. OpenRouter claims it will only actually switch models when the expected benefit of switching exceeds the switching cost, and it will try to stay with the same model as much as possible.


OpenRouter's internal testing claims that across 423 tasks in 4 Agent benchmarks, Jev Router completed 237, while Auto Router completed 130, solving approximately 82% more tasks. In another 5 Agent benchmarks, its median first-token latency was also lower than other routers tested.


Similar approaches have previously appeared in the Jev community. For example, `gholtzap/jev-codex-model-and-effort-router` first has Jev select a model and reasoning tier for Codex, then locks the entire thread to reuse the cache; `Jev-Auto-Router` attempts to re-select the model on a per-call basis. OpenRouter has now built this type of approach directly into its official routing service.

举报 Correction/Report
Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish