We used BE-PACE to evolve new cytosine base editors (CBEs) that overcome target sequence context constraints of canonical CBEs. One evolved CBE, evoAPOBEC1-BE4max, is up to 26-fold more efficient at editing GC, a disfavored context for wild-type APOBEC1 deaminase, while maintaining efficient editing in all other sequence contexts tested. Another evolved deaminase, evoFERNY, is 29% smaller than APOBEC1 and edits efficiently in all tested sequence contexts. We also evolved a CBE based on CDA1 deaminase with much higher editing efficiency at difficult target sites. Finally, we used data from evolved CBEs to illuminate the relationship between deaminase activity, base editing efficiency, editing window width, and byproduct formation. These findings establish a system for rapid evolution of base editors and inform their use and improvement.Genome editing has revolutionized the life sciences and entered clinical trials to treat genetic diseases. 1 The use of programmable nucleases to generate double-stranded DNA breaks (DSBs) followed by homology-directed repair can introduce a wide variety of modifications but is inefficient in non-dividing cells, and is typically accompanied by an excess of unwanted insertions and deletions (indels), translocations, or other chromosomal rearrangements. 2 Base editing directly modifies target DNA bases in living cells and has become widely used to correct or install point mutations in organisms ranging from bacteria to human embryos. 3 Base editors use a catalytically impaired Cas9 to open a single-stranded DNA loop at a specified genomic site (Fig. 1a). Bases within the editing window (typically ~5 nt wide) in this region are modified by a tethered base-modification enzyme that only accepts single-stranded DNA.