Long-context claims still fall apart past 100k tokens

I've been running the same multi-document QA set through a few of the current long-context models, and the gap between the advertised window and the usable window is still huge. Retrieval at 200k looks fine on needle tests,