The big win is multi-file refactors where the model needs to hold the full dependency graph in context simultaneously. With 200k, compaction kicks in mid-task and the model loses track of which files it already changed, leading to inconsistencies or repeated work. With 1M you can load an entire module (tests, types, implementation) and get a coherent refactor in one pass.

Also useful for long debugging sessions where the conversation history itself is the context. Compaction preserves a summary, but summaries lose the exact error messages, stack traces, and failed approaches that help the model avoid going in circles.