Codex CLI extremely slow after billing renewal, one-word prompt takes 2-4 minutes even in empty directory

Open 💬 2 comments Opened Jun 6, 2026 by AlbertHowar

What version of Codex CLI is running?

codex-cli 0.136.0

What subscription do you have?

ChatGPT Pro $100

Which model were you using?

gpt-5.5 high

What platform is your computer?

Linux 6.17.0-PRoot-Distro aarch64 unknown

What terminal emulator and version are you using (if applicable)?

Termux on Android, using proot-distro Ubuntu as root.

Codex doctor report

Additional doctor findings:

codex doctor --json completed and shows a major local-state slowdown:

  • codexVersion: 0.136.0
  • install: npm on linux-aarch64
  • auth: ChatGPT tokens configured
  • MCP servers: 0
  • network/provider reachability: ok
  • websocket handshake: ok
  • current executable: /usr/local/lib/node_modules/@openai/codex/...
  • enabled feature flags include: plugins, apps, browser_use, computer_use, skill_mcp_dependency_install
  • terminal warning: width 59 cols, but this should not explain multi-minute delay

Important:
state.paths took 223618 ms, about 3m43s.

It reports:

  • active rollout files: 23 files
  • total size: 227710372 bytes
  • average size: 9900450 bytes

This suggests the local Codex state/rollout files are very large. However, the strongest signal is still plugins:

Command:
time codex exec --disable plugins --skip-git-repo-check "Answer exactly one word: OK"

Result:
real 0m12.187s
tokens used: 8,387

With plugins enabled, the same one-word prompt takes around 2-4 minutes.

What issue are you seeing?

What issue are you seeing?

Codex CLI became extremely slow after my ChatGPT Pro $100 billing renewal. Before the renewal, standard mode was fast enough, even without "/fast". Now even a one-word prompt in an empty directory takes around 2-4 minutes.

This happens outside my project repo and also with a clean CODEX_HOME test, so repo context, project AGENTS.md, RTK, normal ~/.codex config, and session cache do not appear to be the root cause.

Tested:

  • codex-cli 0.137.0: empty-directory one-word prompt took around 4m20s.
  • codex-cli 0.136.0: same test still took around 2 minutes.
  • Clean CODEX_HOME: same slow behavior.
  • RTK/read-only commands were ruled out separately.
  • The issue still happens in an empty directory with a trivial prompt.

Thread ID from "/feedback":
019e9cd5-3197-7513-9cc8-434cceb190ca

What steps can reproduce the bug?

What steps can reproduce the bug?

Uploaded thread:
019e9cd5-3197-7513-9cc8-434cceb190ca

Steps:

  1. Open Termux on Android.
  2. Enter proot Ubuntu.
  3. Run:

mkdir -p /tmp/codex-empty-test
cd /tmp/codex-empty-test
codex

  1. Prompt Codex with:

Answer exactly one word: OK

  1. Observe that Codex takes around 2-4 minutes to answer a single word.

Also tested with clean Codex home:

mkdir -p /tmp/codex-clean-home
cd /tmp/codex-empty-test
CODEX_HOME=/tmp/codex-clean-home codex

Then prompted:

Answer exactly one word: OK

The slowdown remained.

What is the expected behavior?

What is the expected behavior?

A one-word response in an empty directory should complete within seconds. There is no repo context, no test execution, no file modification, no approval/guardian review, and no meaningful tool work required.

Additional information:

The problem started after ChatGPT Pro $100 billing renewal.

Fast mode was not enabled before either, and standard mode used to be much faster.

I originally noticed the slowdown inside my crypto-autopilot repo, but I reproduced it in an empty directory and with clean CODEX_HOME, so the issue seems unrelated to repo size or project instructions.

RTK was tested separately. Read-only RTK commands no longer trigger approval/guardian review, and the empty-directory test does not depend on RTK.

In one empty-directory test, Codex read "superpowers:using-superpowers SKILL.md", but clean CODEX_HOME did not fix the slowdown, so skill/config cache does not seem to be the only cause.

Additional information

_No response_

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗