OpenRouter benchmarks Jev Router against six other routers
Seven routers, six benchmarks, one weighted index you can re-weight yourself, and no placement numbers in the announcement.
OpenRouter has published side by side benchmarks for seven model routers, Jev Router among them, measured on quality, speed and cost across six benchmarks. Leading individual models are included as comparisons, so the question the page answers is not just which router wins but whether routing beats picking one model and staying there.
Routers aren't always better.
Model switches require rebuilding the input cache, task complexity isn't easily judged from prompts, and selection logic adds latency.
The launch set is split into three kinds of router. Blends of models covers Unbiased Pareto and Sakana Fugu. Model selectors covers OpenRouter's own Auto Router and Jev Router. Model switchers covers NVIDIA Switchyard. Jev Router sits in the selector bucket, meaning it picks a model for the task rather than blending outputs or swapping mid stream.
The index is a weighting, not a verdict
Each benchmark gets a Router Index score weighted 60% quality, 20% time per task and 20% cost. OpenRouter put a slider on the page so you can change those weights, which is the honest move and also the catch. A router that leads at the default weighting can fall behind once you push cost to the front, and nothing in the announcement tells you where any router lands, Jev Router included. If you want a placement figure you have to go to the page and set the weights that match your workload.
OpenRouter also wrote down why routing can lose. Model switches require rebuilding the input cache, task complexity is not easily judged from the prompt alone, and the selection logic itself adds latency. That is the vendor of the benchmark listing the failure modes of the thing it is benchmarking, which is worth more than the usual launch copy.
For people building on it
This is a published eval across six benchmarks, not one person's run, but it is OpenRouter's own harness and OpenRouter's own Auto Router is one of the seven entrants. Read the Jev Router numbers with that in mind.
It is also the second public set of numbers on Jev Router in roughly a week. Theo spent $1,000 running his own benchmark earlier and found it almost 5x slower on DeepSWE. The two are different harnesses with different workloads, so treat them as separate data points rather than a trend, and check whether the time per task component of the index reflects what you saw when you ran it yourself.