Do Newer Models Hold Up as Context Fills? (tiberriver256.github.io)

A recent experiment has tested the performance of various AI models, including gpt-5.1-codex-max, gpt-5.2-codex, and claude-sonnet-4.5, as their context windows fill beyond 50%. The study aimed to challenge common assumptions about model degradation at higher context fill levels, set up using deterministic structured text to create prompts around 10,000 tokens long. The tests evaluated the models' ability to recall three specific facts and generate an SVG image of a pelican riding a bicycle at different context levels (0%, 25%, 50%, 75%, 98%). The findings from this experiment, which was executed using Codex and Copilot CLI tooling, could have significant implications for the AI/ML community, especially for developers relying on these models in complex tasks requiring substantial contextual information. Researchers anticipate that improved performance as context windows expand can lead to more robust applications in real-world scenarios, such as summarization or long-form content generation. This study exemplifies the ongoing efforts to better understand and leverage the capabilities of next-generation AI models in handling extensive context.