RAG vs Long Context in 2026: When to Retrieve and When to Just Stuff the Window

Claude Opus 4.7 ships with a 1 million token context window. Gemini 2.5 has 2 million. GPT-5 sits at 400k. The obvious question: do we still need RAG, or can we just paste the whole codebase into the prompt? After rebuilding two production features both ways, the answer is not what I expected.