Home › Docs › Routing

Routing

apicuts sends every request on the route that gives you the lowest cost for a successful, quality-approved result. Not the cheapest token price: the cheapest task that actually works.

The rule: lowest cost per successful task

A cheaper model that fails, retries or produces worse output costs more than a stronger one that gets it right the first time. So we measure the cost of the whole task, including retries and follow-up calls, and route to whatever is cheapest at the quality you need.

Provider routing

For the model you ask for, apicuts picks among verified, authorized providers using:

• Effective price, including cached-token discounts, not just list price.
• Cache affinity, so requests that share a prompt prefix keep hitting the provider that holds the cache.
• Latency and throughput, so you keep the same speed.
• Error rates and capacity, so traffic avoids providers that are struggling.

Same model, same output quality. Any model or provider you use today can be routed through apicuts.

Automatic load balancing

Traffic is spread across healthy providers and routes automatically, so no single provider becomes a bottleneck. Combined with failover, it keeps the same speed when one provider is busy.

Optional model routing

Some requests don't need your largest model. With your approval, we test lower-cost equivalent models against your own evaluations first, roll them out to a small share of traffic, and only then send simple requests their way. Complex requests stay on the stronger model. Nothing switches models without your sign-off.

Automatic failover

If a provider errors, slows down or goes offline, eligible requests move to a healthy route automatically, so production keeps running. See reliability and status.

Region rules

Choose the EU or the US, and requests only go to providers and regions that match. See security and data residency.

Your routing rules

We set your rules with you during onboarding: which models are allowed, latency budgets, region and data-handling constraints, and which routes may use cheaper models. A self-serve policy editor is on our roadmap.