We carried out a critical appraisal of the two theoretical models, Kurucz' ATLAS9 and PHOENIX/NextGen, for stellar atmosphere synthesis. Our tests relied on the theoretical fit of spectral energy distributions (SED) for a sample of 334 target stars along the whole spectraltype sequence, from the classical optical catalogs of Gunn & Stryker (1983) and Jacoby et al. (1984). The best-fitting physical parameters (T eff , log g) of stars allowed an independent calibration of the temperature and bolometric scale vs. empirical classification parameters (i.e. spectral type and MK luminosity class); in addition, the comparison of the synthetic templates from the ATLAS and NextGen grids allowed us to probe the capability of the models to match spectrophotometric properties of real stars and assess the impact of the different input physics. We can sketch the following main conclusions of our analysis:i) fitting accuracy of both theoretical libraries drastically degrades at low T eff , where both ATLAS and NextGen models still fail to properly account for the contribution of molecular features in the observed SED of K-M stars.ii) Comparing with empirical calibrations, both ATLAS and NextGen fits tend, in average, to predict slightly warmer (by 4-8%) T eff for both giant and dwarf stars of fixed spectral type, but ATLAS provides in general a sensibly better fit (a factor of two lower σ of flux residuals) than NextGen.iii) There is a striking tendency of NextGen to label target stars with an effective temperature and surface gravity in excess with respect to ATLAS. The effect is especially evident for MK I-III objects, where a fraction of stars of about one in four is clearly misclassified by NextGen in log g. This is a consequence of some "degeneracy" in the solution space, partly induced by the different input physics and geometry constraints in the computation of the integrated emerging flux (ATLAS model atmospheres assume standard plane-parallel layers, while NextGen adopts, for low-gravity stars, a spherical-shell geometry). A different T (τ ) vertical structure of stellar atmosphere seems also required for NextGen synthetic SEDs in order to better account for limbdarkening effects in cool stars, as supported by the recent observations of the EROS BLG2000-5 microlensing event.