Reliable uncertainty estimation of battery State Of Health (SOH) is essential for ensuring the safety and stable operation of lithium-ion batteries. Although numerous uncertainty-based SOH prediction methods have been proposed, existing approaches still suffer from overly conservative confidence intervals, limited interpretability, and low computational efficiency. To address these challenges, this work introduces Stable Extended Kalman Filter (EKF) -Transformer, a novel framework that integrates a Transformer module for capturing temporal dependencies with an Extended Kalman Filter module for dynamic uncertainty propagation. Specially, the Transformer module is designed to model the comprehensive temporal dynamics, including both long- and short-term dependencies of the battery degradation process. The Extended Kalman Filter module then employs the Transformer’s output and the predicted priori state to estimate the Jacobian matrix, update the covariance, and refine the state estimation. This process accomplishes dynamic propagation and precise calibration of uncertainty. A key contribution for computational efficiency is our diagonal Jacobian parameterization. This design employs learnable diagonal elements, contributing to both a significant reduction in computational cost and enhanced algorithmic stability. Experimental results on multiple public datasets demonstrate that StableEKF-Transformer achieves superior accuracy, uncertainty quantification, and computational efficiency. It achieves a 1–5% reduction in RMSE and a 0.5–3% reduction in MAE, alongside more reliable and compact confidence intervals, as evidenced by a higher PICP and a lower PINMW. A key efficiency achievement is a 71–92% reduction in parameters and a 46–98% lower computational cost, facilitated by our diagonal Jacobian parameterization, enabling real-time performance and suitability for lightweight vehicle battery management system solutions.