I did a bit of benchmarking using our rspec suite and the most cost effective flavor seems to be "c1.medium". It's almost twice as fast as s1.small but still has the same per-core price. The micro instances were about 4 times slower than c1.medium.
The c1.xlarge instance did not perform better than c1.medium (actually worse) and costs as much per core.
I did not try the cluster nodes as they require you to create your own image to use it.