In this short paper, we study the photometric precision of stellar light curves obtained by the CoRoT satellite in its planet-finding channel, with a particular emphasis on the time scales characteristic of planetary transits. Together with other articles in the same issue of this journal, it forms an attempt to provide the building blocks for a statistical interpretation of the CoRoT planet and eclipsing binary catch to date. After pre-processing the light curves so as to minimise long-term variations and outliers, we measure the scatter of the light curves in the first three CoRoT runs lasting more than 1 month, using an iterative non-linear filter to isolate signal on the time scales of interest. The behaviour of the noise on 2 h time scales is described well by a power-law with index 0.25 in R-magnitude, ranging from 0.1 mmag at R=11.5 to 1 mmag at R=16, which is close to the pre-launch specification, though still a factor 2-3 above the photon noise due to residual jitter noise and hot pixel events. There is evidence of slight degradation in the performance over time. We find clear evidence of enhanced variability on hour time scales (at the level of 0.5 mmag) in stars identified as likely giants from their R magnitude and B-V colour, which represent approximately 60 and 20% of the observed population in the directions of Aquila and Monoceros, respectively. On the other hand, median correlated noise levels over 2 h for dwarf stars are extremely low, reaching 0.05 mmag at the bright end. The CoRoT space mission, launched on December 27, 2006, has been developed and is operated by the CNES, with the contribution of Austria, Belgium, Brazil, ESA, Germany, and Spain. CoRoT data become publicly available one year after release to the Co-Is of the mission from the CoRoT archive: http://idoc-corot.ias.u-psud.fr/.