“While some LLMs are marketed with million-token context lengths, their effective, reliable context window is closer to 200,000 tokens before performance starts to struggle.”