OpenAI quietly trims Codex context window from 372k to 272k tokens

A merged pull request to OpenAI’s Codex repository lowers the advertised context window for its bundled model from 372k to 272k tokens — a roughly 27% cut in how much code and conversation history the coding agent can hold at once. The change wasn’t shipped as a headline capability decision; it landed inside a JSON model-metadata file under a PR titled ‘Backport refreshed bundled model metadata to 0.144,’ framed as a routine hotfix to the 0.144 release branch.

That framing is exactly why it drew attention. Context length is one of the most-watched specs for coding agents, since it governs how many files, diffs, and prior turns the model can reason over before it starts dropping context. Reducing it — even if the intent was to correct metadata to match the model’s real limits rather than to claw back capability — directly affects users working across large codebases. The single-file, low-ceremony nature of the edit meant the reduction surfaced through community scrutiny of the diff rather than any announcement.

The PR itself offers no stated rationale beyond synchronizing bundled metadata with the release, so whether this reflects a genuine capability rollback, a correction of previously overstated numbers, or a backend model swap remains unclear from the change alone. The episode is a reminder that for open-source agent tooling, meaningful product behavior often changes through unannounced config commits that only get noticed when someone reads the file.