Accuracy can mean many things. Smart-seq2, the most reliable technique that is currently available, appears to pick up between 20-40% of the transcripts that are present in a cell (http://www.nature.com/nmeth/journal/v10/n11/full/nmeth.2639.html)
Hi Xiaoxiao, I don't recall that they looked at the error rate, though it may be hidden somewhere deep in the paper. I would assume that the main source of error rate will be in the final high-throughput sequencing, which I doubt will be very different between single cell RNA-seq and normal (population-wide) RNA-seq. Considering that the reads are nowadays quite long, I expect that in most cases this will not compromise the unique mapping.