Background: Ankle-foot exoskeletons can reduce muscle activity and metabolic energy consumption during walking. Experimental studies consistently show that submaximal assistance, typically providing 40–60% of the biological plantarflexion moment, achieves the largest reductions in metabolic power. Inverse musculoskeletal simulations, which neglect factors such as comfort and balance, also predict that submaximal assistance is optimal, yet at higher levels than those observed experimentally. The biomechanical mechanisms underlying this submaximal assistance and its disagreement with experimental observations remain poorly understood. Methods: In this study, we addressed these questions using inverse simulations of walking and systematically examining muscle-tendon dynamics, metabolic power, and ankle joint stiffness across ten assistance levels (0–100% of the biological plantarflexion moment). Results: We found that the bi-articular role of the gastrocnemius limits the reduction of metabolic power at the highest levels of assistance. A sensitivity analysis modifying biceps femoris short head strength provided causal confirmation of this mechanism. We further show that increasing assistance substantially reduces ankle joint stiffness by up to 55%, primarily due to de-recruitment of the soleus. Conclusions: These findings identify both the explanation for submaximal assistance in inverse simulations and propose a missing control objective in exoskeleton simulation frameworks. Because standard effort-based cost functions do not penalize reductions in joint mechanical impedance, this trade-off between effort minimization and stiffness preservation may help explain why simulation-based optimal assistance exceeds experimentally observed values. Inverse simulation does not model robustness, for which impedance preservation may serve as a tractable formulation. Accounting for such a term might represent a new direction for more accurate prediction of optimal exoskeleton assistance.