Highly pathogenic avian influenza A viruses (HPAIV) of the H5N1 subtype occasionally transmit from birds to humans and can cause severe systemic infections in both hosts. PB1-F2 is an alternative translation product of the viral PB1 segment that was initially characterized as a pro-apoptotic mitochondrial viral pathogenicity factor. A full-length PB1-F2 has been present in all human influenza pandemic virus isolates of the 20(th) century, but appears to be lost evolutionarily over time as the new virus establishes itself and circulates in the human host. In contrast, the open reading frame (ORF) for PB1-F2 is exceptionally well-conserved in avian influenza virus isolates. Here we perform a comparative study to show for the first time that PB1-F2 is a pathogenicity determinant for HPAIV (A/Viet Nam/1203/2004, VN1203 (H5N1)) in both mammals and birds. In a mammalian host, the rare N66S polymorphism in PB1-F2 that was previously described to be associated with high lethality of the 1918 influenza A virus showed increased replication and virulence of a recombinant VN1203 H5N1 virus, while deletion of the entire PB1-F2 ORF had negligible effects. Interestingly, the N66S substituted virus efficiently invades the CNS and replicates in the brain of Mx+/+ mice. In ducks deletion of PB1-F2 clearly resulted in delayed onset of clinical symptoms and systemic spreading of virus, while variations at position 66 played only a minor role in pathogenesis. These data implicate PB1-F2 as an important pathogenicity factor in ducks independent of sequence variations at position 66. Our data could explain why PB1-F2 is conserved in avian influenza virus isolates and only impacts pathogenicity in mammals when containing certain amino acid motifs such as the rare N66S polymorphism.