Context. Amongst the new techniques currently developed for high-resolution and high-dynamics imaging, the hypertelescope architecture is very promising for direct imaging of objects such as exoplanets. The performance of this instrument strongly depends on the co-phasing process accuracy. In a previous high-flux experimental study with an eight-telescope array, we successfully implemented a co-phasing system based on the joint use of a genetic algorithm and a sub-aperture piston phase diversity using the object itself as a source for metrology. Aims. To fit the astronomical context, we investigate the impact of photon noise on the co-phasing performance operating our laboratory prototype at low flux. This study provides experimental results on the sensitivity and the dynamics that could be reached for real astrophysical observations. Methods. Simulations were carried out to optimize the critical parameters to be applied in the co-phasing system running in the photon-counting regime. We used these parameters experimentally to acquire images with our temporal hypertelescope test bench for different photon flux levels. A data reduction method allows highly contrasted images to be extracted. Results. The optical path differences have been servo-controlled over one hour with an accuracy of 22.0 nm and 15.7 nm for 200 and 500 photons/frame, respectively. The data reduction greatly improves the signal-to-noise ratio and allows us to experimentally obtain highly contrasted images. The related normalized point spread function is characterized by a 1.1 × 10-4 and 5.4 × 10-5 intensity standard deviation over the dark field (for 15 000 snapshots with 200 and 500 photons/frame, respectively). Conclusions. This laboratory experiment demonstrates the potential of our hypertelescope concept, which could be directly transposed to a space-based telescope array. Assuming eight telescopes with a 30 cm diameter, the I-band limiting magnitude of the main star would be 7.3, allowing imaging of a companion with a 17.3 mag.