GPT-6 automatic context compaction became much slower after recent updates (seconds to 2–3+ minutes, Pro $200/month)

What version of the Codex App are you using (From “About Codex” dialog)?

Currently installed Windows package: OpenAI.Codex version 26.908.4834.0 (verified using Get-AppxPackage, not the About dialog).

The exact last-known-good version and first affected version have not been identified.

What subscription do you have?

ChatGPT Pro, $200/month.

What platform is your computer?

Windows x64 (Microsoft Windows NT 10.0.26200.0).

What issue are you seeing?

I am experiencing a substantial slowdown in automatic context compaction while using GPT-6 in Codex.

A few days ago, automatic compaction was very fast and sometimes completed in just a few seconds. After the last few version updates, it has become extremely slow: some compactions now take 2–3 minutes or longer.

These long pauses interrupt my workflow and are a noticeable regression from my recent experience on the same Pro subscription.

The durations above are my observations during normal use, not timings from a controlled benchmark. I noticed the slowdown after recent updates, but I cannot yet establish whether the cause is a client update, a backend change, or another factor.

What steps can reproduce the bug?

Observed workflow (not yet a deterministic minimal reproduction):

1. Use GPT-6 in the Codex desktop app with a ChatGPT Pro subscription.
2. Continue working until automatic context compaction is triggered.
3. Observe the time taken for compaction to finish and the task to continue.
4. In recent use, some compactions take 2–3 minutes or longer, compared with sometimes only a few seconds a few days earlier.

Exact context sizes and per-event timestamps were not captured for this report.

What is the expected behavior?

Automatic compaction should remain fast and predictable, comparable to the recent experience where it could finish in seconds. Please investigate the substantial increase in latency and restore the earlier responsiveness where possible.

Additional information

• Reported on September 14, 2026 (Asia/Shanghai).
• This report concerns compaction latency specifically.
• The separately installed command-line tool currently reports codex-cli 0.154.0; this is not being presented as the desktop app's bundled engine version or as proof of reproduction in the CLI.
• Please check whether recent client releases or backend changes could explain this regression, and advise which targeted diagnostics would be most useful to identify the cause.

No source file, test, or code entry point is identified. Start by reproducing automatic context compaction on the Windows package OpenAI.Codex 26.908.4834.0 while recording context sizes and per-event timings; done means identifying whether the regression is client-side or backend-related and restoring predictable compaction latency.