My company recently went through the same thought process and ended up using CQL3 over thrift.
Although there is a slight lack of transparency with the additional layer of abstraction going on with CQL3, the ease and familiarity of writing SQL style statements makes the code much more readable and intuitive in my opinion. Plus we found the cqlsh interface far more user friendly than cassandra-cli for debugging and general db maintenance (the auto-complete is fab in cqlsh!).
Once you understand the underlying data structure and how CQL3 represents that data, the extra layer of abstraction pales into insignificance, really.