Theo spent $1,000 benchmarking OpenRouter's Jev Router
The router picks a model per request. Theo's numbers say you could have just picked one yourself.
OpenRouter shipped typesafe/jev-router, a cache-aware router that uses Jev to pick the model and reasoning effort for each request, balancing quality, speed and cost. Theo called it a bad idea the same day, then stayed up until 2am and spent $1,000 running it against a benchmark.
The result is unflattering. On DeepSWE, the router performed roughly the same as GPT-6 Astra on low, cost slightly more, and took almost 5x longer to run. That is one benchmark from one person, but it is the first public cost and latency test of the thing, and the direction is not subtle.
I stayed up til 2am and spent $1,000 benchmarking Jev Router so you don't have to.
Why he expected it to fail
Theo's objection is architectural, not a complaint about this implementation. Jev, he says, is a "system one" model. It categorizes, it does not reason. Deciding how complex a coding task is requires understanding the system the prompt was written in, the tools the agent has, and what the agent has to work around to finish the job. The router sees the prompt and nothing else.
His example is the one everybody has typed at some point. Ask an agent to "port this to Rust" and the router cannot tell whether that is a 100-line TypeScript file or a million-line application. Ask it to fix a bug and it cannot tell whether that is a two-line change or a thread through the ugliest part of the codebase.
He also has a second, more boring argument against the economics. Routing to a dumber model partway through a task does not save much, because cache writes are already a massive share of the cost.
He still likes Jev
Worth separating the tool from the use. Theo says he loves Jev, uses it for a ton of things, and calls it one of his favorite releases of the year. This is the same split he has drawn before, most recently when he argued that using Jev to rank reasoning model outputs is its worst use. The classifier is fine. Asking a classifier to estimate how much thinking a job needs is the part that does not work.
For working engineers, the takeaway is not that routing is dead, it is that prompt-only routing has now been priced. Nobody has yet shown a router that reads your repository and your tool surface before choosing, which is the version Theo's critique leaves room for. Until then, picking a model yourself and keeping the cache warm is the boring baseline that a $1,000 benchmark failed to beat.
In order to know how much thinking a task requires, you have to think. Jev does not "think".