Context size reduction

Canonical version: Context size reduction.

The context-window arms race runs in one direction in the marketing (bigger!), but vendors sometimes quietly go the other way. OpenAI reduced the Codex model's context size from 372k to 272k tokens, and because the Codex CLI is open source, the change surfaced as a GitHub pull request rather than an announcement. It hit the Hacker News front page precisely because shrinking feels like heresy.

The discussion laid out why a smaller window may be the right call:

• Cost and latency — serving a bigger window is expensive and slow, and many sessions never come close to filling it. Cutting the ceiling improves the median experience.
• Quality degrades long before the limit anyway. Users reported keeping sessions under 200k by choice because models get noticeably worse past that; my own experience with context rot in long Claude Code sessions matches, although I feel like it's still ok between 250-400K tokens.
• Compaction quality beats raw capacity. One OpenAI Codex user reported never noticing the limit because compaction just works; another compared it to Fable sessions that run 500k tokens without compacting once. What users actually experience is the compaction strategy, not the window size.

Two side observations from the thread worth keeping. Simon Willison spotted that the same commit added destructive-action guardrails to the system prompt (resolve targets with read-only checks first; never target $HOME or / with recursive destructive commands), which is system-prompt-as-changelog archaeology at its best. And the fact that basic specs like context size are undocumented, discoverable only through X threads and open-source diffs, says a lot about how fast this layer still moves.

The takeaway for anyone building on these models: treat the context window as a budget to manage, not a capacity to fill. See Context Engineering.

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 Window
• Context Engineering
• Agentic Context Management
• OpenAI Codex
• Codex CLI
• Orchestrator tax
• 2026-08-05 OpenAI shrank the Codex context window on purpose

About Sébastien

Ready to get to the next level?