Coalescent-based methods are widely used to infer population size histories, but existing analyses have limited resolution for deep time scales (> 2 million years ago). Here we extend the scope of such inference by re-analysing an ancient peak seen in human and chimpanzee effective population size around 5-7 million years ago, showing that coalescent-based inference can be extended much further into the past than previously thought. This peak is consistently observed across human and chimpanzee populations, but not in gorillas or orangutans. We show that it is unlikely to be an artefact of model violations, and discuss its potential implications for understanding hominin evolutionary history, in particular the human-chimpanzee speciation.