Today, there are many commercial CAM systems available capable of generating NC codes for milling free-form geometries up to 5-axis. However, the quality of these NC codes has not yet been well discussed so far. Apparently, on the computer screen and the geometrical simulation provided by CAM, tool paths calculated by different CAM systems seem to be the same. However, as observed in this work, NC codes differ expressively according to the CAM software applied, and it affects the productivity and quality of the machined surface. Thus, a study was carried out to investigate this issue. A representative workpiece with free-form geometries was designed; a data acquisition system from an open architecture CNC was developed and NC codes were generated by five worldwide commercial CAM systems. The finishing milling operation was evaluated. The results presented difference of up to 30% on the real machining time, differences in the feed rate oscillation and up to 2 times the surface roughness value. This work reveals an essential limitation on the CAM algorithm and arises a new point for benchmarking CAM systems, which brings an opportunity to improve the calculation of tool paths for milling free-form geometries.