Extreme Programming: First Results from a Controlled Case Study "Results shows that while the first release is a learning effort for all stakeholders, the second release shows clear improvement in many regards. The estimation accuracy is improved by 26% and productivity was increased by 12 locs/hour. Yet, the post-release defect rate remained low, i.e., 2.1 defects/KLoc."
Has a nice graph of effort in XP projects here (page 5).
The Costs and Benefits of Pair Programming "Significantly, the resulting code has about 15% fewer defects". Also shows that the same functionality is implemented in up to 25% less code.