OpenAI shrank the Codex context window on purpose

Canonical version: OpenAI shrank the Codex context window on purpose.

While every model announcement brags about bigger context windows, OpenAI went the other way: the Codex model's context size dropped from 372k to 272k tokens. No announcement; the change surfaced as a pull request in the open-source Codex CLI repo, and promptly hit the Hacker News front page because shrinking feels like heresy.

It isn't. Serving a bigger window costs more and responds slower, most sessions never come close to filling it, and model quality degrades long before the hard limit anyway. Experienced users keep sessions short by choice.

The lesson I take from it: Treat AI context as a budget to manage, not a capacity to fill. Spend tokens on what the current task needs, keep durable knowledge in specs and instruction files outside the window, send exploration to subagents, and end long runs deliberately before quality sags. If the vendor building the model thinks a smaller window serves you better, believe them.

Bonus find from the same commit, spotted by Simon Willison: new system-prompt guardrails telling Codex to resolve targets with read-only checks before destructive actions, and never to point a recursive delete at $HOME or /. Open-source harnesses turn system prompts into changelogs.

References

• Hacker News discussion: https://news.ycombinator.com/item?id=48965850
• The PR: https://github.com/openai/codex/pull/33972/files
• Explanation from Thibault Sottiaux: https://x.com/thsottiaux/status/2076543065045795309

Related

• Context size reduction
• Treat AI context as a budget to manage
• Context Engineering

About Sébastien

Ready to get to the next level?