codex ships with a model catalog, and its gpt-5.6 entry lists the context window as 272,000 tokens. the published spec for the model is 1,050,000. 272,000 is also where the api reprices: past that many input tokens the whole request bills at 2x input and 1.5x output, including the tokens under the line. the obvious read is that the window was set to keep sessions on the cheap side of it. that is not the reason openai gave. thibault sottiaux said the driver is "overall cost of cache reads ...