Get early access Open the demo

The cheapest model that can actually do the job.

Most messages do not need a frontier model. Auto-Route works out which ones do, on every single request, and bills you accordingly.

Mechanism

Four steps, on every message.

1

Classify

Every message is scored on task type, required context length, whether tools or web access are needed, and whether the answer depends on multi-step reasoning. This runs on a small local classifier, adds roughly 20ms, and never leaves our infrastructure.

2

Match against measured quality

Each model carries a rolling quality score per task class, built from your workspace's own thumbs-up and thumbs-down history plus aggregate signal across Swurl. A model that keeps getting corrected on legal summarisation stops receiving legal summarisation.

3

Pick the cheapest that clears the bar

Not the best model. The cheapest one whose measured score for this task class is above the threshold you set — conservative, balanced, or aggressive. The difference between those three settings is usually 30% of your bill.

4

Fall back on degradation

Error rates and latency are watched continuously. A provider that breaches its SLO is pulled out of rotation within a minute and traffic shifts to the nearest equivalent. You see a note in the thread rather than a failure.

Evidence

Cost falls, output does not.

What it looks like on a real bill

A 12-seat workspace running roughly 15,000 messages a month. Same work, same people, routing switched on halfway through the period.

Task classShare of messagesRouted toCost per 1k
Classification and extraction31%Fast tier$0.18
Drafting and rewriting24%Fast tier$0.22
Coding19%Balanced tier$1.40
Analysis over long documents17%Frontier tier$4.90
Multi-step reasoning9%Reasoning tier$9.60

Pinning every message to a frontier model would have cost $676. Auto-Route delivered the same work for $284.

You keep the overridePin any thread to a specific model and Auto-Route leaves it alone permanently. Pin at the workspace level and it never runs at all.
Routing decisions are loggedEvery message records which model was chosen and why, visible in the thread and exportable from usage analytics. Nothing about the decision is hidden from you.
It is not magicRouting trades a small amount of ceiling for a large amount of cost. On work where only the best possible answer will do, pin to a frontier model and pay for it.

Turn it on and watch the bill.

Auto-Route is included on every paid plan. Switch it off any time; nothing about your threads changes.

Open the workspace