Based on Hristo's timing, it appears to be a clear win.  

A near-wash for truly string-only dicts that shouldn't be effected; a near-wash for looking up non-(exact-)strings, and a nearly 40% speedup for the target case of looking up but not inserting a non-string or string subclass, then looking up strings thereafter. 

Additional comments:

Barring objections, I will promote from patch review to commit review when I've had a chance to look more closely.  I don't have commit privs, but I think some of the others following this issue do.

The test looks pretty good enough -- good enough that I wonder if I'm missing something on the parts that seem odd.  It would be great if you either cleaned them up or commented to explain why:

Why is the first key vx1, which seems, if anything, like a variable? 
 Why not k1 or string_key?

Why is the first key built up as vx='x'; vx += '1' instead of just k1="x1"?

Using a str subclass in the test is a great idea, and you've created a truly minimal one.  It would probably be good to *also* test with a non-string, like 3 or 42.0.  I can't imagine this affecting things (unless you missed an eager lookdict demotion somewhere), but it would be good to have that path documented against regression.

This seems like a test that could probably be rolled into a bigger testfile for the actual commit.  I don't have the name of such an appropriate file at hand right now, but will try to find it on a deeper review.