Context window

Context windows come up on ToTheMoon not as spec-sheet numbers but as an everyday constraint the hosts hit in their own projects: how many calls, pages or code files fit into one chat before the model starts to forget. The term appears in three storylines — ChatGPT's memory, coding agents such as Claude Code and Codex, and the price of access to the strongest models. The question is always practical: is the window big enough for a real task, and what happens when it overflows?

In 2025 the window was a reason for optimism. Episode 055 treated o3's larger context as part of a workbench where hundreds of pages can be uploaded without switching products. In early 2026, episode 092 offered plain advice: wait for the window to grow, then load everything at once. In episode 102 context became the argument for Google — whoever has enough context to be useful every day wins — and the hosts complained that ChatGPT could not build a window out of their own search history.

The assessment then hardened. Episode 104 noted that the effective context window cannot hold everything needed to reproduce a person's identity. Episode 108 called the improved handling of context an important but overlooked change in ChatGPT 5.5, and episode 119 described how, in a single chat loaded with calls, the window effectively dies and the system starts dropping data, while Claude Code and Codex clean context automatically. Episode 126 observed that the ChatGPT 5.5 Pro window had long exceeded a million tokens, and episode 131 that for a large company no window is enough to solve a task on the first pass.