The stability of the relationship between neural activity and external variables is a matter of debate for many brain areas [1-5]. Whether neural representations are stable or "drifting" has significant implications for how downstream neural populations process these representations and incorporate them into computations driving perception, behavior, and memory. Most studies on the stability of neural coding for movements have focused on simple, stereotyped, and overtrained behaviors, finding that neural representations are stable [6-8]. However, it is unclear if this stability results from the automatic and overtrained nature of the tasks or if it is a general feature of motor coding. Studying behaviors with higher trial-to-trial variance increases our understanding of neural control in more complex naturalistic and unconstrained behaviors. We study representational drift in motor cortical coding using an adapted version of the Whishaw single-pellet reaching task [9-11]. Our version allows for free movement (not head-fixed), self-initiation, and places the pellet atop a narrow pedestal, increasing task difficulty by requiring greater precision in paw placement and spatiotemporally coordinated digit control (dexterity). These adaptations evoke high trial-to-trial variance in forelimb, paw, and digit movements, minimizing automaticity even after the task is well-learned. We use fine-grained behavioral descriptions of paw, digit, and head kinematics in unrestrained mice [12,13] and record the calcium fluorescence activity of large neuronal populations (~300 per field of view) across multiple cortical layers simultaneously, tracking them across days. Our trial-averaged results corroborate previous findings of stability in the motor cortex [6]: peri-event time histograms (PETHs) locked to reach onset are strongly correlated across days for the majority of neurons regardless of laminar position. Using single-trial methods we also find stability. Linear models predicting the activity of individual neurons from kinematic variables show that most cells encode the same kinematic features across days. Kinematic components of the paw and digits, reflecting the dexterous phase of the reach, grasp, and carry, as well as head position, are strong and stable predictors of single-trial activity. These results suggest that rodent motor cortical representations are stable at the single-cell level, even for high-variance naturalistic tasks requiring significant dexterity.