Codex feels extremely slow even in Fast mode; same symptoms as #24431 and #24427

What issue are you seeing?

Codex has become extremely slow for me over the last couple of days. Even when Fast mode is enabled, the actual response speed feels far below what I expect from Fast mode. In some turns it feels like output is not even reaching roughly 5 tokens/second, and the time before useful output appears is long enough to interrupt normal coding work.

This seems similar to the symptoms reported in:

• #24431: GPT-5.5 performance and reliability seem significantly worse today
• #24427: Windows Desktop: many threads in one project make sending and responses very slow even on fast model

I am seeing the same general class of problem: Codex feels much slower than usual, and Fast mode does not feel meaningfully fast. The slowdown affects ordinary coding/debugging workflow, not just one unusually large request.

What steps can reproduce the bug?

1. Open Codex and use it for a normal coding/debugging task.
2. Enable Fast mode.
3. Ask Codex to inspect files, reason about code, or continue an implementation task.
4. Observe that the response/stream startup and actual output speed are extremely slow.
5. Compare the behavior with prior Codex sessions where Fast mode produced visibly faster turnaround.

What is the expected behavior?

Fast mode should have clearly better turn latency and output responsiveness than Standard mode. It should not feel like it is producing only a few tokens per second during normal coding tasks.

What is the actual behavior?

Fast mode currently feels very slow, close to or worse than Standard-mode expectations. This makes interactive coding work difficult because each turn takes much longer than expected, while usage is still consumed as Fast-mode work.

Additional information

I am filing this as an additional user report because my symptoms match #24431 and #24427. I do not want to include private local logs, account details, or workspace paths in the public issue, but I can provide redacted diagnostics if maintainers specify what would be useful.