Models change every month.Your config should not.
RouterLane is an LLM API built for coding agents. You choose how much speed and depth you want. We choose the model that delivers it, task by task.
Why tiers instead of models
An agent session is not one kind of work. In ten minutes it sends a typo fix, a question about a flag, a hard debugging task and a "thanks". A single model name in a config file treats them all the same: too slow and expensive for the small turns, or too weak for the hard one.
So the choice moves to where the information is: the task itself. You pick a lane that says how you want to trade speed for depth. A small decision model reads each new task, and the lane's rules pick the model and the reasoning effort. Inside a task, the decision holds, and the conversation stays on one model so its cache stays warm.
When a better model ships, the lanes change underneath you. The name in your config stays smart.
Principles
Measured, not guessed
Every default comes from a benchmark of graded tasks run at every effort. GLM 5.3 at max effort passed fewer tasks than at high and took far longer, so Smart runs high. Smart+ is measured end to end on real repositories, and the numbers are on its page.
You see every decision
Every response names its tier, model, effort and the source of the decision in x-router-* headers. Nothing about routing is hidden from the client that paid for it.
Your bill is per tier
You pay the rate of the lane you chose, whichever model answered. A hard task that moves up to a stronger model does not change the price of the lane.
Talk to us
Write to hello@routerlane.com, or use the contact form.
Pick a lane.Ship the work.
Keys are issued on request. Tell us what you are building and which clients you use, and we will send a key by email.