The term "world model" is used broadly in AI, encompassing everything from latent dynamics models to traffic scenario generators. This ambiguity led to the development of "State of World Models 2026: Taxonomy, Benchmarks and Open Challenges," aiming to provide a consistent way to describe these models. The report defines a world model as an AI that learns an environment's representation to predict, simulate, evaluate, or support actions within it. This broad definition includes various AI applications but excludes generative models that lack essential environmental consistency.A universal ranking is deemed misleading because different world models excel in distinct areas, such as visual realism, robot planning, or safety testing. Instead, the report proposes a taxonomy based on practical fields like domain, input/output modalities, action conditioning, representation, temporal horizon, and evaluation type. The domain, such as robotics or video generation, significantly influences a model's purpose and evaluation criteria. Functionality is another key differentiator, with models serving purposes like prediction, simulation, planning, or data generation.Internal representations vary from pixels to latent vectors and symbolic variables, each with its trade-offs. The temporal horizon, from next-state prediction to procedural planning, is crucial as errors can accumulate over time. Action conditioning, distinguishing between passive prediction and "what if I do this" scenarios, is a vital practical distinction. Evaluation is fragmented across perceptual, physical, functional, and planning aspects, highlighting the "perception-functionality gap."The report suggests a structured catalog for models and benchmarks to facilitate filtering and comparison. It emphasizes documenting known information, separating evidence from interpretation, and implementing versioning to manage the rapidly evolving field. Exclusions are necessary to maintain focus, preventing the catalog from becoming an all-encompassing AI directory.