In an insurance context, consider {X-n, n > 1} random claim sizes with common distribution function F and {N(t), t >= 0} an integer valued stochastic process that counts the number of claims occurring during the time interval [0, t]. Based on the number of near-extremes which are the observations X-1 near the largest or the mth largest observation, we derive in this article a strongly consistent estimator of upper tails of X-1. Furthermore, estimators for both the tail index and the upper endpoint are introduced when F is a generalized Pareto distribution. Asymptotic normal law for the proposed estimators is additionally presented.