This Codex change is easy to misread if you already use GPT or Claude and you are starting to chain AI tools together just to save yourself hassle. The trap is putting Codex in the same bucket as every other model and picking by the bigger number. When you are bouncing from a browser to a chat box to an editor, then repeating the same background again, a bad read here means more copying context by hand and one more round of rework.
The headline is clear: "OpenAI reduces Codex Model Context Size from 372k to 272k." My read is just as clear: 272K is first a billing line, not an intelligence line. AI tools are not just competing for code work. They are competing for the scraps of time you lose switching between windows.
The reason is in the numbers. The GPT-5.6 Sol model page still lists a 1,050,000-token context window, which just means how much text the model can hold in one go. But that same page also says that once input goes past 272K, the whole request is billed at 2x input and 1.5x output [S001]. That is not a tiny footnote. That is the point where "can fit" and "should send" stop meaning the same thing.
Then there is the Codex pricing change itself. As of April 2, 2026, Codex pricing is framed around input, cached input, and output tokens [S005]. That does not prove why OpenAI made the cut, and I would not pretend it does. What it does support is the practical interpretation: 272K is a clean economic boundary, so routine work should be designed around it before anyone starts telling a dramatic "the model got dumber" story.
A lot of people think they need a smarter model. What they really need is fewer window swaps. My default move would be simple: keep normal Codex tasks at or under 272K, and only go past it when the extra context is worth the price jump. If you know someone setting default context limits by headline numbers alone, share this with them. The real question is not "did Codex get weaker?" It is "where does the bill change, and will that actually save me work?"