OpenAI cut Codex context from 372K to 272K

OpenAI merged a pull request on July 19 that cut Codex's effective GPT-5.6 context window from 372K tokens to 272K, per the file diff at openai/codex#33972. That number isn't random. OpenAI prices any prompt above 272K input tokens at 2x input and 1.5x output. The multiplier applies to the whole request, not just the overage. The cap and the price break are the same number.

No blog post. No deprecation notice. Just a number changed in model metadata, caught by a few people before Hacker News piled 356 points onto the thread.

One engineer reported burning about $400 in an afternoon. Their harness kept feeding repo context past the new boundary, still assuming the old 372K limit. Cursor and Windsurf don't surface the change either, so anyone running Codex through a wrapper is likely eating errors or 2x charges without knowing why.

I run my own setup that feeds a lot of repo context to models. That made me check whether anything I own assumes a published limit instead of measuring one. Two places did, both reading a constant I'd copied from documentation eight months ago.

What to actually do

Stop treating vendor-published context limits as stable config. Read them from the API when you can. When you can't, log actual input token counts per call and alert when they trend toward the ceiling you believe in. An error you notice immediately. A bill you notice on the first of the month.

Worth connecting: an independent tracker at codex-resets.com logs every Codex usage reset. It hit Hacker News the same day, showing 35 resets so far, averaging 8.9 days apart. Somebody built monitoring for another company's quota policy because the alternative was finding out the hard way.