Cassandra's counters don't support the 'increment and get' semantics of general atomic counter implementations. So what you are asking is impossible with Cassandra's counters. The reason is that it would not be possible without coordination between the replicas to ensure values returned are always the latest.
To do this, you will need a distributed system with atomic semantics such as ZooKeeper. A thread that will solve your problem on ZooKeeper is here: http://www.mail-archive.com/zookeeper-user@hadoop.apache.org/msg01967.html.