You send a tier as the model name. The router reads each new task, then picks a model and a reasoning effort inside that tier. The same tier makes different calls for a typo and a deadlock, and that is the point.
Each cell is the model and effort a tier uses for a task of that complexity. The kind of task is recorded on every decision, and rules can match on it too.
| Tier | trivial | simple | moderate | hard |
|---|---|---|---|---|
| Fast | DeepSeek 4.1 Flashoff | DeepSeek 4.1 Flashoff | DeepSeek 4.1 Flashmedium | DeepSeek 4.1 Flashmedium |
| Normal | DeepSeek 4.1 Flashoff | DeepSeek 4.1 Flashmedium | DeepSeek 4.1 Flashmedium | GLM 5.3high |
| Smart | DeepSeek 4.1 Flashlow | GLM 5.3medium | GLM 5.3high | GLM 5.3high |
| Smart+ | A job on RouterLane's machineexplore and QA: DeepSeek 4.1 Flash, high. build: GLM 5.3, high | |||
Max effort is on the scale and no default rule uses it: on GLM 5.3 it passed 85% of the benchmark against 100% at high, with a p90 of 71.6 s against 20.5 s.
fastQuestions, lookups, small edits and the quick back-and-forth inside a session. Thinks only when the task needs it. Close calls round down.
normalEveryday coding: features, fixes, tests and reviews, with enough reasoning to get it right the first time. Close calls go to the nearest level.
smartHard debugging, design, plans and multi-step changes. The strongest model and deep reasoning. Close calls round up.
smart-plusWhole jobs on a repository: a feature with tests, a migration, a refactor across files. The work runs on RouterLane's machine and one patch comes back.
Sample prompts, decided by each tier's default rules.
A mechanical edit. No thinking time, so the first byte arrives fastest.
One fact, one answer. Fast replies straight away.
Real investigation. Fast stays on its quickest model and turns on medium reasoning.
Chit-chat gets no thinking time.
Everyday feature work: the balanced default.
Hard debugging moves up to the larger model at high effort.
Smart still answers a thank-you quickly.
A small task on the strongest model, with medium reasoning.
Design work gets the strongest model at high effort. When the call is close, Smart rounds up.
Explore and QA on DeepSeek 4.1 Flash, build on GLM 5.3. One upload, one patch back.
The project's own tests run on the copy before the patch is sent.
The client declared no shell tool, so nothing can be uploaded. The turn runs as a Smart turn.
A new task is read by the router's classifier, a small decision model that answers two questions in 200 to 350 ms. The rest comes from the request itself.
trivial, simple, moderate or hard. Judged on the work, not the length of the message: a short "yes, do it" inherits the difficulty of the request it answers.
chat, question, edit, feature, debug, refactor, plan, review, ops or writing. Recorded on every decision and available to rules.
When the classifier is unsure, the tier decides which way to round: Smart up, Fast down, Normal to the nearest level.
How many, and how many tokens they add. Large images move a request from GLM 5.3 to GLM 5.3 Vision.
The estimated length of the request. Past 85% of a model's window, it moves to the 1M-token model.
Whether this is a tool result inside a running task, which model the conversation is on, and whether the new task is harder than the last.
Continuation turns skip the classifier entirely. Everything else is read once, then cached for 24 hours by task.
When the newest message is a tool result, the request belongs to the task already running. The router reuses that task's model and effort: no classification and no delay. The response says x-router-source: continuation.
A conversation keeps its model, so the upstream prompt cache stays warm: 5K to 22K cached input tokens per turn in real agent loops. When a later task is harder than the one that set the model, the conversation moves up to the stronger one.
When a request passes 85% of the chosen model's window, it moves to DeepSeek 4.1 Flash, which reads 1,048,576 tokens. Long sessions keep going instead of failing on length.
Every tier takes images. When images push a request on GLM 5.3 past about 8K tokens, it moves to GLM 5.3 Vision.
| Choose a model yourself | Choose a tier | |
|---|---|---|
| Who picks per call | You, once, in code | The router, per task |
| A "thanks" in a hard session | Billed and answered like the hard work | Read as trivial and answered without deep reasoning |
| Reasoning effort | One setting for everything | Set per task, from off to high |
| Prompt cache | Lost when you switch models by hand | Kept: the conversation stays on one model |
| A model errors or stalls | Your retry code | Failover before the first byte |
| A better model ships | You benchmark it and redeploy | The tier name in your code stays the same |
| Price | Changes with the model | One rate per tier |
| Seeing what happened | Logging you build | x-router-* headers on every response |
Keys are issued on request. Tell us what you are building and which clients you use, and we will send a key by email.