Balanced cantilever prestressed concrete box-girder bridges worldwide are known to exhibit ongoing and excessive deflections due to creep and shrinkage, which are often underestimated by conventional code-based creep and shrinkage models. These models also exhibit substantial uncertainties, leading to wide spreads in long-term structural behavior predictions and raising concerns regarding serviceability, durability, and structural safety. This study presents a computationally efficient Bayesian updating framework that integrates short- and long-term deflection measurements to improve the predictive performance of creep and shrinkage models. The framework employs Variational Bayesian Monte Carlo (VBMC) for sample-efficient posterior approximation and couples it with a zoned beam element model to accurately capture the effects of cross-sectional variability while remaining computationally efficient for direct likelihood evaluation. Application to the Rooyensteinse Bridge demonstrates that prior to updating, both fib Model Code 2020 (MC20) and RILEM B4 (B4) underestimate midspan deflections (errors of 65 % and 20.6 %, respectively) and exhibit high predictive uncertainty. Bayesian updating substantially reduces prediction errors to 4.3 % for MC20 and 1.7 % for B4, while decreasing the associated relative predictive uncertainties by 86 % and 82 %, respectively. The updated models also show that long-term prestress losses increase along the full length of the bridge, rising from 0.8 % to 7.5 % for MC20 and from 4.0 % to 19.7 % for B4 after 45 years at the hammerhead. Drying shrinkage is identified as the most underestimated parameter, with updated values increasing by factors of 3.5 (MC20) and 4.9 (B4). The perceived importance of drying shrinkage strongly depends on the number and duration of incorporated deflection measurements. Measurements from the first decade of service life were insufficient to capture the governing mechanisms responsible for the accelerated long-term deflection trend, resulting in an underestimation of the multi-deflection deflections. Although B4 still provides plausible multi-decade predictions due to its broader uncertainty bounds and flexible parameterization, MC20 becomes overconfident. Incorporating multi-decade measurements improves the accuracy of both models, allowing them to reliably reproduce the observed deflection trend. These results highlight the importance of well-structured models and the integration of long-term measurement data for reliable creep and shrinkage predictions in balanced cantilever box-girder bridges.