“The performance of large context window models like Google's Gemini degrades significantly as the context window approaches its limit, for example at 800,000 tokens out of a 1 million token capacity.”