The thermal softening related to the stress relaxation, i.e. a reduction in the internal resistance to deformation, due to the increasing bulk temperature is theoretically studied using the incubation time approach. The plastic deformation at high-loading rates is often accompanied by an obvious adiabatic temperature rise. Phenomenological constitutive models describing these thermo-mechanical responses of metals are generally developed by assuming empirical rate-dependent and temperature-dependent components into existing conventional bases originally proposed for quasi-static cases. This work aims to model the thermal softening of metals subjected to high-rate loads using the incubation time approach and examine the temperature–time correspondence, i.e. temperature-dependence of the incubation time, by introducing the relative stress (RS) factor over a wide range of strain, strain rate and temperature. The incubation time approach considers the strain-rate sensitivity as a manifestation of the time sensitivity of materials. Additionally, we conduct a comprehensive analysis of the developed relaxation model of plasticity (RP model) and conclude that this model may be derived by the time-dependent yield surface equation using the incubation time framework. Based on experimental facts for HSLA–65 steel, 93W–4.9Ni–2.1Fe Tungsten-based composite and titanium alloy Ti–6Al–4V, the descriptive abilities of the developed RP model are compared with other constitutive models (phenomenological and micromechanism-based ones) and the artificial neural networks (ANN) model. The advantages and disadvantages of various models are discussed, and the main differences between the ANN model and other constitutive models are examined.