Machine learning models on tabular datasets must adapt efficiently to evolving feature spaces—a common scenario in domains such as healthcare and insurance, where new diagnostic tests or policyholder attributes are introduced over time. Existing solutions either require full retraining, which is computationally expensive, or rely on transformer-based or variational autoencoder (VAE) architectures that incur significant computation overhead. We propose a cost-effective model update framework that incorporates Feature-Incremental G-Shapley (FIGS) to enable fast adaptation to newly added features. FIGS estimates changes in data Shapley values to identify samples whose marginal contributions to model performance shift most after feature expansion. These informative points are selected as a coreset for fine-tuning, enabling accurate model updates using only a small subset of data. Unlike prior Shapley value estimations that assume static feature spaces and require costly recomputation, FIGS efficiently predicts contribution shifts from differences in feature representations. Experiments on medium- and large-scale tabular datasets, across multiple backbone models, show that FIGS consistently outperforms existing coreset selection strategies. Our proposed framework achieves near full-retrain accuracy with significantly reduced update time. The implementation repository is open-sourced at: https://github.com/tzralice/FIGS .