To understand the cortical control of complex limb movements, we must identify how movement control signals differ across different sensorimotor regions. Here we characterize the structure of movement-related activity across 6 subregions of primary and secondary motor (M1 and M2) and primary somatosensory cortex (S1) in the mouse. To evoke coordinated forelimb movements, we developed a reach-to-grasp task in which head-fixed mice performed delayed reaches to 15 distinct target locations. We performed two-photon calcium imaging in layer 2/3 of each area (>8,000 neurons, 19 sessions, 4 mice). These recordings revealed a clear anterior-to-posterior gradient across the cortex in single-cell and population-level activity features. Single cells were most sharply tuned in M2: many cells (26\%) were modulated by movements to $\le 5$ targets, with 55\% modulated by $\ge$ 8 targets. In contrast, <1\% of M1 and S1 cells were modulated by $\le 5$ targets while 87\% were modulated by $\ge 8$ targets. Neural population trajectories in M2 were most distinct across targets, with separation during the delay consistent with movement preparation. Separation occurred later for M1 and S1, consistent with a primary role in movement execution and the arrival of somatosensory feedback. Target-specific activity was also more persistent in M2: classifier dimensions fit around movement onset remained performant later in time. In contrast, M1 and S1 classifiers generalized less well to other time points, arguing for activity features that more closely tracked ongoing movement details instead of target identity. These detailed movement representations supported better joint angle decoding from M1 and S1 (73\% proximal, 35\% distal variance explained in M1; 67\%, 38\% in S1) than from M2 populations (45\%, 21\%). Overall, these results suggest that anterior regions more strongly reflect abstract movement selection, while posterior regions reflect the kinematic and sensory details of movement execution.