Codex quietly caps GPT-5.6 context at 272K
A July 18 metadata backport locked Codex's effective context window to OpenAI's 272K pricing threshold, down from the 372K it previously reported.
A pull request merged into the Codex CLI on July 18 rewrote the bundled model metadata for GPT-5.6 Sol, Terra, and Luna, and in the process cut their reported context window from 372,000 tokens to exactly 272,000. The prior figure combined a 372K window with a 95% effective-context multiplier, landing usable context around 353,400 tokens in practice. The new number removes that multiplier and hard-caps the window at 272,000 instead.
That figure isn’t arbitrary. OpenAI prices any prompt over 272,000 input tokens at double the input rate and 1.5x the output rate for the rest of that session, across standard, batch, and flex tiers. By capping Codex’s default window at exactly that line, OpenAI stops users from drifting into the expensive tier by accident, but it also means Codex users can no longer reach the longer context window GPT-5.6 is otherwise advertised as supporting, 1.05 million tokens, without leaving Codex entirely. Developers on GitHub have been filing this as a regression rather than a fix; one thread is titled “SEVERE REGRESSION” and tracks the window shrinking twice in recent weeks.
If your workflow depends on Codex holding a large repo or long conversation in context, budget for the 272K ceiling now, and if you genuinely need the full window, that means going around Codex to the API directly rather than assuming a config flag will bring it back. It’s one more entry in a pattern worth tracking alongside the Codex economics paper and the Codex API’s own July 23 sunset: the product keeps changing shape under long-running workflows built on top of it.