272K Is a Pricing Line, Not an Intelligence Line

你一会儿在浏览器搜资料，一会儿回聊天框补背景，一会儿再回编辑器改几行代码。

最容易做错的，是把 Codex 当成同一种工具，以为谁分高谁就适合自己。；代价往往是如果把它们都当成同一种工具，你会在最该省事的地方继续手动搬运上下文，最后多一轮返工。；我先给一个保守判断：272K首先是计费线，不是智力线。

The real cost shows up in a boring place: browser, chat, editor, back to browser, back to chat, back to editor. AI tools are not just coming for coding work. They are coming for the fragments of time you lose switching between windows.

That is why my read on "OpenAI reduces Codex Model Context Size from 372k to 272k" is conservative but practical: 272K is a pricing line before it is an intelligence line.

The GPT-5.6 Sol model page still lists a 1,050,000-token context window, but it also says requests above 272K are billed at 2x input and 1.5x output. Then OpenAI's April 2, 2026 Codex billing update split costs into input, cached input, and output against a shared team AI usage balance. Put those two together and 272K reads first as an economic boundary, not automatic proof the model got worse. Boundary: this read comes from the docs and billing update, not from latency benchmarks or community reports.

A lot of people think they need a smarter model. What they really need is fewer windows.

So the next decision is not "which number is bigger?" It is "where should we cap default context so we stop paying to paste background over and over?

真正该讨论的是：AI 工具真正开始抢的，不只是代码活，而是你来回切换的那些碎时间。