The dispersion of gases in complex situations such as the case of buildings in close proximity is a difficult problem, but important for the safety of people living and working in such areas. Computational fluid dynamics (CFD) provides a method to build and run models that can simulate gas dispersion in such geometrically complex situations; however, the accuracy of the results needs to be assessed. As a first step in such an assessment, this study considers the simulation of the dynamics of the basic atmospheric boundary layer using the FLUENT CFD code and the prediction of gas dispersion from a single stack. The CFD results are compared with the predictions from the Atmospheric Dispersion Modelling System (ADMS), a well tested and validated quasi-Gaussian model. When FLUENT was set up to simulate the neutrally stable atmospheric boundary layer, the mean velocity profiles were well predicted and were maintained with downwind distance. The algebraic Reynolds stress turbulence model provided the best predictions for the turbulence kinetic energy (TKE) and dissipation. The dissipation rate was maintained throughout the length of the model domain and, on average, the TKE levels were within 80% of the expected values up to a height of 100 m, but at the ground reduced to 50% of the inlet values. Predictions of TKE using the simpler k–ε model turbulence was much poorer. Spread of the gas plume were predicted using an advection-diffusion (AD) method, a Lagrangian particle tracking (LP) method and a large eddy simulation (LES) method. The LP method gave the best results; the horizontal and vertical plume spreads were similar to those predicted by ADMS and ground level and plume centre line concentrations were close to ADMS values. However, some differences were observed with the ground level concentrations rising more rapidly with distance than for ADMS, but reaching similar peak values while the plume centreline concentrations dropped more rapidly than in ADMS. For the AD method the horizontal cross-wind plume spread was significantly lower than expected resulting in higher ground level concentrations than predicted by ADMS, an effect that was attributed to the isotropic formulation of the AD equation in FLUENT. The LES results were intermediate between the AD and LP predictions. Overall, the CFD simulations with the LP method were satisfactory; however, they could not be considered as an appropriate alternative to a model such as ADMS for normal atmospheric dispersion studies because of the much larger run times and the greater complexity of setting up model runs. CFD is more appropriate for applications that involve complex geometry that could not be simulated using ADMS; however, further studies are required to assess the ability of CFD to calculate dispersion in such situations, for instance, around groups of buildings and under a range of atmospheric stability conditions, rather than just the neutral stability considered in this paper.