Recent multimodal large language models (MLLMs) are credited with three capabilities: they fuse vision and text, forecast the consequences of an action, and recover from mistakes through reflection. A fourth assumption underpins every agentic deployment built on these models: that strong single-step performance transfers to competent multi-step behavior. Existing evaluations cannot adjudicate these claims cleanly: static perception suites never let small errors compound, and agent suites aggregate perception, planning, and control failures into a single score. No instrument exposes where, along the perceive-forecast-act-reflect-recover loop, a multimodal model fails. We introduce Cube Bench, a controlled diagnostic built on the Rubik’s Cube, where transitions are deterministic and oracle distance to the goal is available at every state. Seven oracle-scored tests isolate each stage of the decision loop under matched conditions. Across open and closed-source MLLMs, current models do not reliably exhibit any of the three credited capabilities, and the fourth assumption fails most sharply: a model with frontier-adjacent single-step skill across perception, prediction, and control can sit at the floor on recovery from its own errors. This dissociation between local skill and multi-step recovery has direct implications for agentic systems that assume these capabilities compose.