Visual information in natural scenes is distributed over a broad range of intensities and contrasts. This distribution has to be compressed in the retina to match the dynamic range of retinal neurons. In this study we examined how cones perform this compression and investigated which physiological processes contribute to this operation. M- and L-cones of the goldfish were stimulated with a natural time series of intensities (NTSI) and their responses were recorded. The NTSI displays an intensity distribution which is skewed towards the lower intensities and has a long tail into the high intensity region. Cones transform this skewed distribution into a more symmetrical one. The voltage responses of the goldfish cones were compared to those of a linear filter and a non-linear biophysical model of the photoreceptor. The results show that the linear filter under-represents contrasts at low intensities compared to the actual cone whereas the non-linear biophysical model performs well over the whole intensity range used. Quantitative analysis of the two approaches indicates that the non-linear biophysical model can capture 91 +/- 5% of the coherence rate (a biased measure of information rate) of the actual cone, where the linear filter only reaches 48 +/- 8%. These results demonstrate that cone photoreceptors transform an NTSI in a non-linear fashion. The comparison between current clamp and voltage clamp recordings and analysis of the behaviour of the biophysical model indicates that both the calcium feedback loop in the outer segment and the hydrolysis of cGMP are the major components that introduce the specific non-linear response properties found in the goldfish cones.