Context : A set of 20 extremely red galaxies at 2.5 \leq z _ { phot . } \leq 3.8 with photometric features of old passive-evolving galaxies without dust , with stellar masses of \sim 10 ^ { 11 } M _ { \odot } , have colors that could be related to passive-evolving galaxies with mean ages larger than 1 Gyr .
This suggests they have been formed , on average , when the Universe was very young ( < 1 Gyr ) .

Aims : We provide new estimates for the stellar content of these 20 galaxies , with a deeper analysis for two of them that includes spectroscopy .

Methods : We obtained , with the GRANTECAN-10.4 m , ultraviolet rest-frame spectra of two galaxies and analyzed them together with photometric data .
The remaining 18 galaxies are analyzed only with photometry .
We fit the data with models of a single-burst stellar population ( SSP ) , combinations of two SSPs , as well as with extended star formation .

Results : Fits based on one SSP do not provide consistent results for the blue and red wavelengths .
Moreover , the absence in the spectra of a break at \sim 2 \times 10 ^ { 3 } Å indicates that a rather young component is necessary .
Using two SSPs we can match the photometric and spectroscopic data , with the bulk of the stellar population being very old ( several Gyr ) and the remaining contribution ( < 5 % of stellar mass fraction ) from a young , likely residual star formation component with age \lesssim 0.1 Gyr .
Exponentially decaying extended star formation ( \tau ) models improve slightly the fits with respect to the single burst model , but they are considerably worse than the two SSP based fits , further supporting the residual star formation scenario .

Conclusions : The fact that one SSP can not match these early-type galaxies highlights the limitations for the use of age estimators based on single lines or breaks , such as the Balmer break used in cosmic chronometers , thus questioning this approach for cosmological purposes .