Context. The third Gaia Data Release covers 34 months of data and includes the second Gaia catalogue of long-period variables (LPVs), with G variability amplitudes larger than 0.1 mag (5–95% quantile range).

Aims. The paper describes the production and content of the second Gaia catalogue of LPVs and the methods we used to compute the published variability parameters and identify C-star candidates.

Methods. We applied various filtering criteria to minimise contamination from variable star types other than LPVs. The period and amplitude of the detected variability were derived from model fits to the G -band light curve wherever possible. C stars were identified using their molecular signature in the low-resolution RP spectra.

Results. The catalogue contains 1 720 558 LPV candidates, including 392 240 stars with published periods (ranging from 35 to ∼1000 days) and 546 468 stars classified as C-star candidates. Comparison with literature data (OGLE and ASAS-SN) leads to an estimated completeness of 80%. The recovery rate is about 90% for the most regular stars (typically miras) and 60% for SRVs and irregular stars. At the same time, the number of known LPVs is increased by a factor of 6 with respect to literature data for amplitudes larger than 0.1 mag in G , and the contamination is estimated to be below 2%. Our C-star classification, based on solid theoretical arguments, is consistent with spectroscopically identified C stars in the literature. Caution must be taken in crowded regions, however, where the signal-ro-noise ratio of the RP spectra can become very low, or if the source is reddened by some kind of extinction. The quality and potential of the catalogue are illustrated by presenting and discussing LPVs in the solar neighbourhood, in globular clusters, and in galaxies of the Local Group.

Conclusions. This is the largest all-sky LPVs catalogue to date. The photometric depth reaches G = 20 mag. This is a unique dataset for research into the late stages of stellar evolution.