CAP only talks about the uncommon unhappy path (given a network partition, what is tradeoff between availability and consistency). PACELC theorem builds on CAP to describe that even in the absence of a network partition, there is a trade-off between latency and consistency (in other words, describing the tradeoffs associated with BOTH the common happy path and uncommon unhappy path).
Azure Cosmos DB offers 5 well-defined consistency models for you to choose from, so that you can choose the right tradeoffs for a given application or scenario. This way, you aren't stuck choosing between the hard extremes of Strong and Eventual consistency.
As the Cosmos DB SLA doc describes, for any of the 5 consistency models the service supports, the service guarantees all the other three guarantees (latency at the 99th percentile, availability and throughput) at 99.99%. One of the benefits of well-defined consistency models is that for a given model, developers can clearly make the tradeoffs between (1) latency vs. consistency, (2) availability vs. consistency and (3) throughput vs. consistency. The service documentation covers some of these tradeoffs.
