A systematic assessment of various universal machine-learning interatomic potentials (uMLIPs) is presented, evaluating four different graph neural network (GNN)-based uMLIPs: M3GNet-DIRECT, CHGNet, MACE-MP-0, and ALIGNN-FF. These models are based on GNN architectures that demonstrate transferability across different chemical systems. The evaluation uses data from a recent verification study of density-functional-theory (DFT) implementations and the Materials Project. The goal is to provide guidance for materials scientists in selecting suitable models for their research, offer recommendations for model selection and optimization, and stimulate discussion on potential areas for improvement in current machine-learning methodologies in materials science. The study assesses the performance of these uMLIPs in various calculations, including the equation of state, formation energies, and phonon properties. The results show that MACE performs best in predicting formation energies and vibrational properties, while CHGNet and M3GNet excel in relaxed geometry predictions. MACE and CHGNet also show superior performance in predicting formation energies across the periodic table, especially for systems containing transition metals, chalcogens, and halogens. M3GNet exhibits relatively high errors, leading to a lower R² value in predicting formation energies, but it still performs well in predicting volume and lattice parameters. ALIGNN shows relatively good results in predicting formation energies but is the most problematic model in geometry optimization and predicting vibrational properties. The study highlights the need for further optimization and training of uMLIPs to fully exploit their capabilities across a broader range of applications. The choice of a particular uMLIP should consider a balance between accuracy and computational efficiency. Future work should focus on improving the performance of these potentials, particularly in areas where current uMLIPs exhibit limitations, such as more accurate prediction of forces and stresses or the capability of learning Born effective charges and electronic dielectric tensors crucial for the vibrational properties of polar materials. The results indicate that currently available uMLIPs can predict ab initio vibrational properties with a typical error of 3.71 meV in the best-case scenario.A systematic assessment of various universal machine-learning interatomic potentials (uMLIPs) is presented, evaluating four different graph neural network (GNN)-based uMLIPs: M3GNet-DIRECT, CHGNet, MACE-MP-0, and ALIGNN-FF. These models are based on GNN architectures that demonstrate transferability across different chemical systems. The evaluation uses data from a recent verification study of density-functional-theory (DFT) implementations and the Materials Project. The goal is to provide guidance for materials scientists in selecting suitable models for their research, offer recommendations for model selection and optimization, and stimulate discussion on potential areas for improvement in current machine-learning methodologies in materials science. The study assesses the performance of these uMLIPs in various calculations, including the equation of state, formation energies, and phonon properties. The results show that MACE performs best in predicting formation energies and vibrational properties, while CHGNet and M3GNet excel in relaxed geometry predictions. MACE and CHGNet also show superior performance in predicting formation energies across the periodic table, especially for systems containing transition metals, chalcogens, and halogens. M3GNet exhibits relatively high errors, leading to a lower R² value in predicting formation energies, but it still performs well in predicting volume and lattice parameters. ALIGNN shows relatively good results in predicting formation energies but is the most problematic model in geometry optimization and predicting vibrational properties. The study highlights the need for further optimization and training of uMLIPs to fully exploit their capabilities across a broader range of applications. The choice of a particular uMLIP should consider a balance between accuracy and computational efficiency. Future work should focus on improving the performance of these potentials, particularly in areas where current uMLIPs exhibit limitations, such as more accurate prediction of forces and stresses or the capability of learning Born effective charges and electronic dielectric tensors crucial for the vibrational properties of polar materials. The results indicate that currently available uMLIPs can predict ab initio vibrational properties with a typical error of 3.71 meV in the best-case scenario.