Here goes my test for your suggestion:

Benchmark: timing 3 iterations of 20.two_string, 30.two_string_lookahe
+ad...
20.two_string: 236 wallclock secs (234.31 usr +  1.62 sys = 235.93 CPU
+) @  0.01/s (n=3)
30.two_string_lookahead: 178 wallclock secs (176.23 usr +  1.52 sys = 
+177.75 CPU) @  0.02/s (n=3)
[download]

So it's better, but still not linear ...