I whish to run a semi-parametric Cox proportional hazard model. I know there are "frailty models" to account for unobserved heterogeneity... to what extent are they different from multilevel models?
Yes, multivariate frailty models generally correspond to multilevel models (most closely to a random-effects model with random intercepts). Individuals within a cluster have the same random effect that impacts the hazard multiplicatively. Usually these random effects are gamma or inverse gaussian distributed.