Radar coincidence imaging (RCI) is a super-resolution staring technique based on the innovative idea of random radiation and wavefront random modulation. To reconstruct the target, sparsity-driven methods are commonly used in RCI, while the prior knowledge of the imaging model requires to be known accurately. However, model error generally exists, which induces the inaccuracy of the model and defocuses the image. We focus on sparsity-driven RCI in the presence of gain–phase error and propose an autocalibration expansion–compression variance-component (AC-ExCoV)-based autofocusing method in a sparse Bayesian learning framework. The algorithm determines the gain–phase error as a part of the RCI process by reconstructing the target and compensating the gain–phase error iteratively. To probabilistically formulate the target reconstruction problem, a probabilistic model is utilized to fully exploit the sparse prior, and then solved using ExCoV. Meanwhile, the gain–phase error is estimated and calibrated to obtain a high-resolution focused image. The AC-ExCoV algorithm demands no prior knowledge about the sparsity or measurement-noise level with significant superiority in computational complexity. Simulation results show that the proposed algorithm obtains a well-focused target image with high reconstruction accuracy.