With this paper, we release accurate photometric redshifts for 1692 counterparts to Chandra sources in the central square degree of the COSMOS field. The availability of a large training set of spectroscopic redshifts that extends to faint magnitudes enabled photometric redshifts comparable to the highest quality results presently available for normal galaxies. We demonstrate that morphologically extended, faint X-ray sources without optical variability are more accurately described by a library of normal galaxies (corrected for emission lines) than by AGN-dominated templates, even if these sources have AGN-like X-ray luminosities. Preselecting the library on the bases of the source properties allowed us to reach an accuracy sigma_(Delta z/(1+z_spec)) \sim0.015 with a fraction of outliers of 5.8% for the entire Chandra-COSMOS sample. In addition, we release revised photometric redshifts for the 1735 optical counterparts of the XMM-detected sources over the entire 2 sq. deg.of COSMOS. For 248 sources, our updated photometric redshift differs from the previous release by Delta z>0.2. These changes are predominantly due to the inclusion of newly available deep H-band photometry H_AB=24 mag. We illustrate once again the importance of a spectroscopic training sample and how an assumption about the nature of a source together with the number and the depth of the available bands influence the accuracy of the photometric redshifts determined for AGN. These considerations should be kept in mind when defining the observational strategies of upcoming large surveys targeting AGN, such as eROSITA at X-ray energies and ASKAP/EMU in the radio band.

We present photometric redshifts and spectral energy distribution (SED) classifications for a sample of 1542 optically identified sources detected with XMM in the COSMOS field. Our template fitting classifies 46 sources as stars and 464 as non-active galaxies, while the remaining 1032 require templates with an AGN contribution. High accuracy in the derived photometric redshifts was accomplished as the result of 1) photometry in up to 30 bands with high significance detections, 2) a new set of SED templates including 18 hybrids covering the far-UV to mid-infrared, which have been constructed by the combination of AGN and non-active galaxies templates, and 3) multi-epoch observations that have been used to correct for variability (most important for type 1 AGN). The reliability of the photometric redshifts is evaluated using the sub-sample of 442 sources with measured spectroscopic redshifts. We achieved an accuracy of $\sigma_{\Delta z/(1+z_{spec})} = 0.014$ for i$_{AB}^*<$22.5 ($\sigma_{\Delta z/(1+z_{spec})} \sim0.015$ for i$_{AB}^*<$24.5). The high accuracies were accomplished for both type 2 (where the SED is often dominated by the host galaxy) and type 1 AGN and QSOs out to $z=4.5$. The number of outliers is a large improvement over previous photometric redshift estimates for X-ray selected sources (4.0% and 4.8% outliers for i$_{AB}^*<$22.5 and i$_{AB}^*<$24.5, respectively). We show that the intermediate band photometry is vital to achieving accurate photometric redshifts for AGN, whereas the broad SED coverage provided by mid infrared (Spitzer/IRAC) bands is important to reduce the number of outliers for normal galaxies.