Accurate estimation of lithium-ion battery (LIB) state of health (SOH) underpins safe operation, predictive maintenance, and lifetime-aware energy management. Despite recent advances in machine learning (ML), systematic benchmarking across heterogeneous real-world cells remains limited, often confounded by data leakage and inconsistent validation. Here, we establish a leakage-averse, cross-battery evaluation framework encompassing 32 commercial LIBs (B5–B56) spanning diverse cycling histories and temperatures (≈4 °C, 24 °C, 43 °C). Models ranging from classical regressors to ensemble trees and deep sequence architectures were assessed under blocked 5-fold GroupKFold splits using RMSE, MAE, R² with confidence intervals, and inference latency. The results reveal distinct stratification among model families. Sequence-based architectures—CNN–LSTM, GRU, and LSTM—consistently achieved the highest accuracy (mean RMSE ≈ 0.006; per-cell R² up to 0.996), demonstrating strong generalization across regimes. Gradient-boosted ensembles such as LightGBM and CatBoost delivered competitive mid-tier accuracy (RMSE ≈ 0.012–0.015) yet unrivaled computational efficiency (≈0.001–0.003 ms), confirming their suitability for embedded applications. Transformer-based hybrids underperformed, while approximately one-third of cells exhibited elevated errors linked to noise or regime shifts, underscoring the necessity of rigorous evaluation design. Collectively, these findings establish clear deployment guidelines: CNN–LSTM and GRU are recommended where robustness and accuracy are paramount (cloud and edge analytics), while LightGBM and CatBoost offer optimal latency–efficiency trade-offs for embedded controllers. Beyond model choice, the study highlights data curation and leakage-averse validation as critical enablers for transferable and reliable SOH estimation. This benchmarking framework provides a robust foundation for future integration of ML models into real-world battery management systems.