← Back to feed

Pull request #33972 in the openai/codex repository — titled “Backport refreshed bundled model metadata to 0.144” — shipped a 100,000-token reduction to Codex’s context window. The change went in quietly. The developer community noticed loudly.

The reduction takes Codex from 372,000 to 272,000 tokens, a 27% cut. For users managing large codebases, reviewing multiple files simultaneously, or running sessions that reference heavy documentation, the difference is immediate: context compaction triggers earlier and more frequently.

What Changed in Practice

Codex uses context compaction to extend sessions beyond the model’s window — summarising earlier context to make room for new turns. The problem is lossy. A compaction event drops resolution on prior reasoning, earlier file states, and design decisions made in the first half of a session.

At 372k, developers reported reaching roughly 40% context utilisation before compaction kicked in during typical large-codebase work. At 272k, that threshold drops back toward 12-20% — the level users had before the larger window arrived. One developer summarised it: “372 was not perfect, but it was so much better and a godsend.”

The frustration is structural: Codex is positioned as the tool for extended, autonomous development work. A 100k context cut is in direct tension with that pitch.

The Timing

The reduction arrived as OpenAI was managing two simultaneous pressures. GPT-5.6 Sol launched in July to enormous demand — OpenAI temporarily removed the 5-hour usage limit for Plus and Pro subscribers to cope with the surge. At the same time, OpenAI merged the ChatGPT desktop app with Codex and Work, repositioning Codex as one mode within a unified ChatGPT product.

Whether the context cut is a cost measure, a capacity response to GPT-5.6 demand, or a metadata correction that had downstream effects on the window the model actually uses is not yet clarified by OpenAI.

The Competitive Context

Claude Fable 5 supports 200,000 tokens with no equivalent compaction cliff. Cursor Composer 2.5, which benchmarks against Codex on coding tasks, uses long-context handling as a differentiator. The 372k window had been one of Codex’s stronger technical arguments against both.

At 272k, Codex lands at a context level that matches or trails several competitors at the same price tier. OpenAI has not published a timeline for restoring the larger window or explained the specific rationale in the pull request.