Signals from different oligonucleotide probes against the same target show great variation in intensities. However, detection of differences along a sequence e.g. to reveal intron/exon architecture, transcription boundary as well as simple absent/present calls depends on comparisons between different probes. It is therefore of great interest to correct for the variation between probes. Much of this variation is sequence dependent. We demonstrate that a thermodynamic model for hybridization of either DNA or RNA to a DNA microarray, which takes the sequence-dependent probe affinities into account significantly reduces the signal fluctuation between probes targeting the same gene transcript. For a test set of tightly tiled yeast genes, the model reduces the variance by up to a factor ∼1/3. As a consequence of this reduction, the model is shown to yield a more accurate determination of transcription start sites for a subset of yeast genes. In another application, we identify present/absent calls for probes hybridized to the sequenced Escherichia coli strain O157:H7 EDL933. The model improves the correct calls from 85 to 95% relative to raw intensity measures. The model thus makes applications which depend on comparisons between probes aimed at different sections of the same target more reliable.