Accurate knowledge of species colour discrimination is fundamental to explain colour based behaviours and the evolution of colour patterns. We tested how the receptor noise limited model, widely used in behavioural ecology, matched actual colour discrimination thresholds obtained using behavioural tests. Guppies (Poecilia reticulata) were first trained to push a target coloured disk placed among eight grey disks of various luminances on a grey plate. Guppies were then tested to find target disks, which varied in colour contrast from the plate. The target disks followed a gradient going from high contrast to inconspicuous against the grey background. We plotted the percentage of correct choices of each colour in the gradient against the model prediction and determined the discrimination thresholds using the inflection point of the fitted sigmoid curve. We performed the experiment on six colour gradients: red, orange, yellow, green, blue and purple. Four colour gradients: red, orange, green and blue, showed a discrimination threshold that matched the model predictions. However, deviations of the model for the yellow and purple gradients suggest that ecological relevance of some colours could affect decision-making in behavioural tests and that we can no longer assume that the rules for colour discrimination are independent of colours.