Book a demo Start building

Routing that knows your traffic.

Send each request to the model that fits it best, and fall back on its own when a provider slows down.

When a provider slows down.

Your rules pick a model

Cheapest model that meets your quality bar, or the fastest one for live chat. You decide per route.

A provider slows down

Strand watches every provider’s latency and errors, second by second.

The request moves on its own

It goes to the next model on your list. Your users never see an error.

  • 01 Your rules pick a model
  • 02 A provider slows down
  • 03 The request moves on its own

A few lines of config.

Each route says what matters: speed, cost or quality. Strand does the rest, and you can change it without a deploy.

routes:
  support-chat:
    goal: fastest
    models: [kite-mini, lyra-3, atlas-70b]
    budget: $40/day
  summaries:
    goal: cheapest
    quality: ">= 0.86"
    fallback: wren-8b

What routing changes.

Latency, −63%

p95 falls as routing learns which model answers fastest for each route.

Spend, −41%

Cheaper models take the easy requests. The big ones only see the hard ones.

Quality, +14%

Scores stay flat or rise, because a release that gets worse never ships.

  • 01 Latency
  • 02 Spend
  • 03 Quality

Questions.

Do I have to change my code?

No. Point your existing client at the Strand endpoint and keep your calls as they are.

Which models can I route to?

Every major hosted provider, open-weight models we run for you, and your own fine-tunes.

What happens when every model is slow?

Strand returns the fastest answer it can within your time limit, and tells you which model gave it.

Where does my data go?

Only to the models you allow, in the regions you choose. Nothing is kept for training.

Works with the rest.

Routing gets better with evaluations and tracing switched on.

Ship your first model this afternoon.