Using budding yeast, we have studied Rad51-dependent break-induced replication (BIR), where the invading 3' end of a site-specific double-strand break (DSB) and a donor template share 108 bp of homology that can be easily altered. BIR still occurs about 10% as often when every 6th base is mismatched as with a perfectly matched donor. Here we explore the tolerance of mismatches in more detail, by examining donor templates that each carry 10 mismatches, each with different spatial arrangements. Although 2 of the 6 arrangements we tested were nearly as efficient as the evenly-spaced reference, 4 were significantly less efficient. A donor with all 10 mismatches clustered at the 3' invading end of the DSB was not impaired compared to arrangements where mismatches were clustered at the 5' end. Our data suggest that the efficiency of strand invasion is principally dictated by thermodynamic considerations, i.e., by the total number of base pairs that can be formed; but mismatch position-specific effects are also important. We also addressed an apparent difference between in vitro and in vivo strand exchange assays, where in vitro studies had suggested that at a single contiguous stretch of 8 consecutive bases was needed to be paired for stable strand pairing, while in vivo assays using 108-bp substrates found significant recombination even when every 6th base was mismatched. Now, using substrates of either 90 or 108 nt-the latter being the size of the in vivo templates-we find that in vitro D-loop results are very similar to the in vivo results. However, there are still notable differences between in vivo and in vitro assays that are especially evident with unevenly-distributed mismatches. Mismatches in the donor template are incorporated into the BIR product in a strongly polar fashion up to ~40 nucleotides from the 3' end. Mismatch incorporation depends on the 3'→ 5' proofreading exonuclease activity of DNA polymerase δ, with little contribution from Msh2/Mlh1 mismatch repair proteins, or from Rad1-Rad10 flap nuclease or the Mph1 helicase. Surprisingly, the probability of a mismatch 27 nt from the 3' end being replaced by donor sequence was the same whether the preceding 26 nucleotides were mismatched every 6th base or fully homologous. These data suggest that DNA polymerase δ "chews back" the 3' end of the invading strand without any mismatch-dependent cues from the strand invasion structure. However, there appears to be an alternative way to incorporate a mismatch at the first base at the 3' end of the donor.