With 1M token context windows, does RAG still make sense?

Gemini and Claude now support 1M+ token contexts. Our first reaction was "does this make our entire RAG pipeline obsolete?" In practice we've found long-context is not free — cost scales linearly, attention degrades on needle-in-haystack tasks, and you can't update a context window as cheaply as updating an index. But for some queries it clearly wins. How are teams thinking about the split between long-context-first vs. retrieval-first?