(abridged) We take advantage of the deep Chandra and Spitzer coverage of a large area (more than 10 times the area covered by the Chandra deep fields, CDFs in the COSMOS field, to extend the search of highly obscured, Compton-thick active nuclei to higher luminosity. These sources have low surface density and large samples can be provided only through large area surveys, like the COSMOS survey. We analyze the X-ray properties of COSMOS MIPS sources with 24$\mu$m fluxes higher than 550$\mu$Jy. For the MIPS sources not directly detected in the Chandra images we produce stacked images in soft and hard X-rays bands. To estimate the fraction of Compton-thick AGN in the MIPS source population we compare the observed stacked count rates and hardness ratios to those predicted by detailed Monte Carlo simulations including both obscured AGN and star-forming galaxies. The density of lower luminosity Compton-thick AGN (logL(2-10keV)=43.5-44) at z=0.7--1.2 is $(3.7\pm1.1) \times10^{-5}$ Mpc$^{-3}$, corresponding to $\sim67%$ of that of X-ray selected AGN. The comparison between the fraction of infrared selected, Compton thick AGN to the X-ray selected, unobscured and moderately obscured AGN at high and low luminosity suggests that Compton-thick AGN follow a luminosity dependence similar to that discovered for Compton-thin AGN, becoming relatively rarer at high luminosities. We estimate that the fraction of AGN (unobscured, moderately obscured and Compton thick) to the total MIPS source population is $49\pm10%$, a value significantly higher than that previously estimated at similar 24$\mu$m fluxes. We discuss how our findings can constrain AGN feedback models.