Anthropic just made the 1M token context window generally available for Claude Opus 4.6 and Sonnet 4.6, and yeah—Google and OpenAI have had “1M context” too. The reason this drop is actually interesting is not the number, it’s the pricing move and the retrieval numbers. Anthropic removed the long-context pricing premium entirely, so 900k tokens is billed the same way as 9k tokens. That’s a huge deviation from the tiered long-context pricing other frontier labs use, and it means you can realistically shove in way more data—think ~600 PDF page images instead of ~100—without getting punished for it. The second piece is retrieval accuracy at scale, using MRCR v2 (8 needles in a haystack), which is closer to real workloads than the old single-needle tests. At 256k tokens, the new Claude models are basically state-of-the-art (around ~90% retrieval), and when you push to 1M tokens, a lot of models fall off a cliff. Gemini drops hard, GPT-5.4 drops too, but Claude retains performance with only ~18% reduction for Opus 4.6—still usable. Practically, this means fewer compactions (agents “forgetting” due to summarization), better multi-round agent performance, and more viable long-running agentic tasks. And no, this doesn’t kill RAG. Docs can exceed 1M tokens, long context increases latency, and if you don’t need to send everything, retrieval is still cheaper and more practical—just likely with more hybrid retrieval strategies than pure embeddings.