Delamination is a dominant and often design-limiting failure mechanism in laminated composites, and its reliable numerical prediction is central to structural certification and virtual testing. Cohesive zone modelling (CZM) provides a practical framework to capture delamination initiation and growth under pure and mixed-mode loading via traction–separation laws, but high-fidelity simulations can be computationally expensive when fine discretisations are required to resolve the fracture process zone and evolving damage. This paper introduces a novel, physics-based acceleration framework for quasi-static mixed-mode delamination analyses in composite laminates. The technique constructs a Reduced Order Model (ROM) by statically condensing Degrees Of Freedom (DOFs) that are not affected by fracture, thus substantially reducing the model size, while preserving the accuracy and generality of the original model. In order to allow for arbitrary delamination propagation, condensation is performed dynamically, with bespoke matrix formats being employed to efficiently handle the resulting operations. Furthermore, it can be combined with different discretisation schemes such as standard or extended finite elements (FEM/XFEM), allowing for increased flexibility. By construction, the accelerated formulation preserves the governing equilibrium and energy balance of the original problem, and is able to reproduce the full-order model exactly, rather than providing an approximate surrogate. The framework is compatible with standard cohesive-interface discretisations and mixed-mode coupling laws, and updates seamlessly as the delamination front evolves. The proposed approach is tested on benchmark laminate delamination configurations spanning pure and mixed-mode I/II loading. Validation is performed by comparing load–displacement response and delamination growth trends to corresponding full-order CZM simulations and to response characteristics reported in the literature. Across all cases, the accelerated simulations replicate the full-order results while delivering substantial reductions in solution time and total solution unknowns - DOFs. The results show that evolving mixed-mode delamination can be accelerated exactly while matching established benchmarks. By drastically reducing the cost per forward solve, the approach is well-suited for workflows such as inverse parameter identification where many forward solutions are needed for calculating fracture parameters