This paper proposes RDO-Gaussian, an end-to-end rate-distortion optimized 3D Gaussian representation that achieves flexible and continuous rate control. The method addresses two main issues in existing 3D Gaussian representations: 1) optimizing rate and distortion simultaneously through dynamic pruning and entropy-constrained vector quantization (ECVQ), and 2) modeling different materials and regions with learnable parameters rather than treating colors equally. RDO-Gaussian significantly reduces the size of 3D Gaussian representations by over 40× and outperforms existing methods in rate-distortion performance. The method is built upon the 3D Gaussian Splatting (3DGS) framework and includes three key components: Gaussian pruning, adaptive spherical harmonics (SH) pruning, and ECVQ. Gaussian pruning removes redundant Gaussians using learned masks, while adaptive SH pruning allows different Gaussians to have different SH degrees. ECVQ quantizes covariance and color parameters to achieve a compact representation. The method is validated on both real and synthetic scenes, demonstrating superior performance in terms of compression ratio and rendering quality. The source code is available at https://github.com/USTC-IMCL/RDO-Gaussian.This paper proposes RDO-Gaussian, an end-to-end rate-distortion optimized 3D Gaussian representation that achieves flexible and continuous rate control. The method addresses two main issues in existing 3D Gaussian representations: 1) optimizing rate and distortion simultaneously through dynamic pruning and entropy-constrained vector quantization (ECVQ), and 2) modeling different materials and regions with learnable parameters rather than treating colors equally. RDO-Gaussian significantly reduces the size of 3D Gaussian representations by over 40× and outperforms existing methods in rate-distortion performance. The method is built upon the 3D Gaussian Splatting (3DGS) framework and includes three key components: Gaussian pruning, adaptive spherical harmonics (SH) pruning, and ECVQ. Gaussian pruning removes redundant Gaussians using learned masks, while adaptive SH pruning allows different Gaussians to have different SH degrees. ECVQ quantizes covariance and color parameters to achieve a compact representation. The method is validated on both real and synthetic scenes, demonstrating superior performance in terms of compression ratio and rendering quality. The source code is available at https://github.com/USTC-IMCL/RDO-Gaussian.