OpenAI cuts the bundled Codex model context window from 372k to 272k

Sources: Codex PR #33972, HN discussion
Summary: OpenAI merged Codex pull request #33972, "Backport refreshed bundled model metadata to 0.144", on 2026-07-18, which lowers the bundled model metadata's context_window and max_context_window from 372000 to 272000 tokens. The change ships in the Codex CLI's bundled metadata for its GPT-5 based model, so the usable context drops by 100000 tokens, about 27 percent, with no model swap.
Why it matters: A cut to the usable context window changes how much code and session history Codex can hold at once, which affects large-repository work and long autonomous agent runs.