Skip to content
Model routing

The right model for every message

You never pick a model. Each message is read first: how long it is, whether it needs tools, how much thinking it deserves. Then it goes to the cheapest model that will do it well.

  • Fast, Balanced or Deep, chosen per message in milliseconds
  • Never a worse answer to save money, only a cheaper route to the same one
  • Moves on if a provider is slow or down
Dana Brooks

@ruby when does the refund window close?

Read as: short question, one lookup

router
  • Fast$0.03 · 0.6 s
  • Balanced$0.12 · 1.8 s
  • Deep$0.48 · 6.2 s

Ruby answered in 0.6 s for $0.03, 94% less than Deep

How it works

  1. 1

    The message is read

    Length, whether tools are needed, how much reasoning the answer takes. This costs a fraction of a cent and a few milliseconds.

  2. 2

    A tier is picked

    A one-line lookup goes to a fast model. A chart with three data pulls to a mid one. A rewrite that needs judgement to the best available.

  3. 3

    Prices are checked live

    Providers change prices and go down. The router reads what things cost now and steps around anything that is not answering.

Questions about this

Can I force a particular model?

No, and that is deliberate. Picking models by hand is how budgets disappear. The router's job is to make that decision better than a person would, every time.

Does the cheap model ever ruin an answer?

The tier is chosen from what the message needs, not from what is left in the budget. A hard message gets the deep model even on the free plan.

How much does it actually save?

On a normal support or ops channel, most messages are short lookups. Teams see the same work for a fraction of what one top model for everything would cost.