OpenAI Cuts Codex Context Window from 372k to 272k Tokens

OpenAI trimmed the Codex model's context window by 100k tokens, dropping the ceiling from 372k to 272k. The change is documented in the pull request on the openai/codex repository.

At 272k tokens, the window still covers large codebases in a single pass for most practical use cases. But engineers who built workflows around the higher limit — feeding in entire monorepo slices, large dependency graphs, or extended conversation histories alongside code — will need to reconsider chunking strategies.

The reduction likely reflects infrastructure cost or stability considerations rather than a capability regression. Large context windows carry significant compute overhead per request, and keeping them at the upper bound imposes pressure on latency and resource allocation. A 100k reduction is meaningful operationally even if the absolute window remains large.

For solo founders and small teams running automated code review, refactoring pipelines, or agent loops against large files, the practical impact depends on how close existing payloads were running to the old ceiling. Workflows comfortably under 272k are unaffected. Those that relied on the extra headroom will require explicit prompt restructuring or retrieval-augmented patterns to compensate.

No public explanation accompanies the change in the pull request diff itself. The team did not annotate the reduction with a migration path or deprecation timeline, which puts the discovery burden on developers monitoring the repository directly.

If you rely on Codex in production or in automated pipelines, audit your maximum payload sizes against the new limit now. Hitting the context ceiling silently or with an opaque error is a common failure mode that surfaces at the worst time — mid-pipeline, not during local testing.

Source