The accurate estimation of soil bearing capacity is fundamental to the safe and economical design of shallow foundations. Conventional analytical and empirical methods often struggle with the complex, non-linear nature of soil behavior and inherent parameter variability. This study investigates the application of Artificial Neural Networks (ANNs) for predicting the ultimate bearing capacity (qult) of shallow foundations on cohesive-frictional soils. A dataset comprising 90 instances with soil unit weight (γ), cohesion (c), angle of internal friction (ϕ, degrees), and friction angle in radians as inputs, and qult as output was compiled. A Multi-Layer Perceptron (MLP) ANN with a 3–10–1 architecture was developed using MATLAB’s Neural Network Toolbox™ and trained using the Levenberg–Marquardt algorithm on \(70\boldsymbol{\%}\) (63 samples) of the normalized data. Model performance was evaluated using standard statistical metrics on the remaining \(30\%\) (27 samples) testing data. The developed model demonstrated excellent predictive capability, achieving correlation coefficients (R) of 0.9994 (Training) and 0.955 (Testing), and low normalized Root Mean Squared Errors (RMSE) of 0.0064 (Training) and 0.0459 (Testing). Regression and scatter plots visually confirmed the strong correlation. A comparative reliability analysis, evaluating the reliability index (β) against a fixed normalized demand, further validated the model’s ability to replicate the statistical characteristics (mean and variance) of the actual data. The results indicate that ANNs provide a robust and accurate alternative for predicting soil-bearing capacity, effectively handling non-linearities and offering a valuable tool for geotechnical engineers.