When Codex feels slow or heavy, the quickest way to find the fix is to first figure out where the time is going. If the AI's replies are slow, check the server and settings. If startup is slow, check MCP servers and the settings folder. If the app screen is heavy, check the app version and conversation length. If running commands is slow, check where your files are stored and your computer's performance. This article lists the fixes by symptom, based on OpenAI's official documentation and the answers OpenAI staff gave on Codex's GitHub (confirmed October 25, 2026).
Decide where to look based on the symptom
Symptom A
The AI's replies are slow or stall
Outages or congestion, reasoning effort, model, thread length → Chapter 2
Symptom B
Startup is slow or you wait before you can type
Waiting for MCP servers to start, the settings folder, an old version → Chapter 3
Symptom C
The app is heavy or freezes
App version, long threads, piled-up runs, Windows issues → Chapter 4
Codex works in the ChatGPT desktop app (Windows and macOS), the CLI (the version you use in a terminal), and editor extensions such as VS Code. In this article, a "thread" means one bundle of work exchanges: one chat in the app, or one session in the CLI. In every form, the model that does the thinking runs on OpenAI's servers, and the commands the AI requests run on your computer. Symptoms are color-coded: response startup app command.
1. Narrow it down first
When something feels slow, look at your screen and check these three things in order.
- Is it waiting for approval?Codex may be stuck waiting for your permission to run a command or change a file. OpenAI's troubleshooting guide also lists this as the first thing to check when Codex appears to be stuck.
- What is it doing right now?If you're waiting before you can type, or the MCP server preparation message stays on screen, that's startup. If "thinking" or "writing a reply" stays on screen, that's response. If a command stays in progress and doesn't move, that's command. If typing or switching screens itself lags, that's an app problem.
- Is it slow for everyone?Check OpenAI's status page for Codex outages. If there is an outage, there is usually little you can do locally.
In the CLI, /status shows the model you're using, the tokens you've used, and the remaining context (the free space in the thread's memory). If the remaining context is low, go to "Don't keep using a long thread" in Chapter 2. If you've set reasoning effort to High or above, go to the table in Chapter 2.
2. When the AI's replies are slow response
Congestion or outages on the server side
Some slowdowns have no local cause. In a report that Fast mode was as slow as Standard, an OpenAI staff member said, "We had a capacity issue last week," and later explained that they had eased the problem (openai/codex #24422, June 2026). This is an example of server-side capacity shortages causing slowdowns. For the message "model is at capacity," see Fixing Codex's "Selected model is at capacity" message.
Match reasoning effort to the task
In Codex, you can choose how long the model thinks before it answers. The stronger the setting, the deeper it thinks, and the more time and usage it consumes. OpenAI's official description is as follows (OpenAI "Models").
| Level | Official description | Suited to (examples) |
|---|---|---|
| Light (Low in the CLI) | For quick, well-scoped work | A fix in one place, answering a question |
| Medium | A balance of speed and depth | Everyday implementation |
| High and Extra High | For difficult work with many steps or materials (OpenAI describes the two the same way)Extra High thinks for longer | Changes that involve design, bugs with unknown causes |
| Max | Prioritizes depth over speed and usage, thinking for longer | Hard problems that couldn't be solved otherwise |
| Ultra | Speeds up complex work by splitting it across subagentsNo official information on the effect on usage | Large tasks that can be divided into several parts |
In the CLI, you choose the level together with the model using /model. You can set the default with model_reasoning_effort in the config file (config.toml). The lightest level is shown as Light in the app and as low in the CLI and the config file.
If you always use High or above, even simple tasks take longer. For small fixes, lowering the setting to Light (low in the CLI) or Medium gets faster replies. OpenAI's model selection page recommends pairing the lighter model GPT-6 Luna with Light (Low) for fine-grained edits and well-scoped work (OpenAI "Model selection").
Fast mode (Fast and Ultrafast)
If you want to cut waiting time, there are faster modes that use more of your usage allowance in exchange for speed (OpenAI "Speed").
| Mode | Target and who can use it | Speed | Usage consumed (when signed in with ChatGPT) |
|---|---|---|---|
| Fast | Supported models, including GPT-6.1 Sol, GPT-6 Astra, GPT-6 Sol, and GPT-6 LunaThe app, CLI, and editor extensions when signed in with ChatGPT | Multiplier not stated officially | 2.5 times the standard plan usagePurchased credits count 2 times |
| Ultrafast | GPT-6 Astra and GPT-6.1 SolPro (the $500/month tier) and eligible Enterprise and Edu plans (pricing, confirmed October 25) | With GPT-6 Astra, text generation is up to 8 times faster than standard | 8 times the usageCredits count 6 times |
In the CLI, switch with /fast. If the model you're using has no faster tier, /fast won't appear. If you sign in with an API key, usage is billed at API rates and the multipliers above don't apply.
Don't keep using a long thread
If you keep working on many tasks in one thread, the context grows, and both replies and the screen get heavier. In a report that threads slowed down after long use, an OpenAI staff member recommended creating a new thread for each task.
OpenAI staff member (openai/codex #11984, February 25, 2026)
"We recommend creating a new thread for each task"
(Creating a new thread for each task is recommended)
The staff member added that when the same thread keeps going, the context gets diluted with each compaction, and that starting fresh is faster and keeps the model from getting distracted by earlier work (openai/codex #11984).
If you really want to keep the same thread, the CLI's /compact summarizes the exchanges so far and frees up context (the official description is "Summarize the visible chat to free tokens"). /new starts a new thread (OpenAI "Slash commands").
3. When startup is slow (CLI and MCP) startup
Waiting for MCP servers to start
If you've registered MCP servers (a way to connect external tools) with Codex, it waits for them to get ready every time it starts. Optional servers are waited for by default for only 1 second before moving on. Servers set to required = true are waited for until they start (up to 10 seconds by default). If you keep registered a server that responds slowly or one you no longer use, the first reply will be delayed. OpenAI's official settings are as follows (OpenAI "Model Context Protocol").
required = true are waited for until they start (startup_timeout_sec, 10 seconds by default). If one can't start, Codex itself won't start.Removing servers that don't need to be required is the most effective fix. Optional servers are waited for only up to mcp_optional_startup_grace_ms (1000 milliseconds by default).enabled = false in that server's section in config.tomlYou can disable a server without deleting its settings.startup_timeout_sec (10 seconds by default)If you lower it too much, servers that take a while to start won't be usable./mcp in the CLI, or codex mcp list in a terminal, to see the enabled servers# Example in ~/.codex/config.toml: stop a server you're not using
[mcp_servers.old-server]
command = "npx"
args = ["-y", "old-mcp-server"]
enabled = falseThe settings folder is large or corrupted
In a report that the CLI started slowly, an OpenAI staff member pointed out that the ~/.codex folder, where settings and records are kept, might be large or corrupted, and suggested temporarily renaming it to test (openai/codex #9290, January 2026). Renaming the folder makes Codex start in the same state as first-time use, so you'll need to sign in again. Past thread records are also kept in this folder, so only rename it, don't delete it, and put it back after testing.
# macOS and Linux: rename the folder
mv ~/.codex ~/.codex-backup
# Windows (PowerShell): rename the folder
Rename-Item $HOME\.codex .codex-backupTo restore, first rename the new .codex that was created during testing to another name (such as .codex-test), then rename .codex-backup back to .codex.
Check whether you're on an old version
Some CLI versions include startup improvements. For example, version 0.152.0 (September 1, 2026) reduced the delay before the first reply, and version 0.161.0 (October 7) detects a corrupted records database earlier, at startup (GitHub releases). Check your version with codex --version, then check for and install updates with codex update. If you installed with a method that codex update doesn't support (such as a package manager), update using the same method you used to install. The Codex included in the desktop app can have a different version from the CLI, and the app is updated separately (see Chapter 4).
4. When the app is heavy or freezes app
Update to the latest version and restart
Many sluggishness problems in the desktop app have been fixed in recent versions.
Windows
In a report that the app freezes or stutters on Windows (openai/codex #20214), an OpenAI staff member wrote on August 28, 2026, that they had fixed one of the causes.
OpenAI staff member (openai/codex #20214, August 28, 2026)
"ChatGPT v26.825.32147 includes a fix for one cause of Windows stuttering and slowdowns during long sessions."
(ChatGPT v26.825.32147 includes a fix for one cause of Windows stuttering and slowdowns during long sessions)
According to the staff member, one cause was commands running in the sandbox piling up resources used by dwm.exe, the Windows component that draws the screen. They advise updating to v26.825.32147 or later and restarting the app.
The staff member says "one cause," and as of October 25, 2026, this issue is still open, with users continuing to report freezes from other causes even after updating. The Windows app is distributed through the Microsoft Store, so check for updates in the Store's Library (OpenAI "ChatGPT desktop app for Windows").
macOS
For the issue of high CPU usage on macOS, a staff member wrote that it was fixed in the April 2026 version (v26.415.20818) (openai/codex #16231). Meanwhile, reports that input lags because helper processes pile up are still unresolved as of October (#25744). Keep the app up to date, and restart it if it becomes heavy.
On either OS, the first step is to update to the latest version and restart the app.
The official steps for when it freezes
OpenAI's troubleshooting guide gives these steps for when things get stuck (OpenAI "Troubleshooting").
git status in the terminal → Start a new thread with a small, narrowly scoped requestpwd or git statusIf it still freezes, wait until the running thread finishes, then restart the app.Clean up threads
Working for a long time in one thread (Chapter 2) also affects how heavy the screen feels. As for the number of threads, there are open GitHub reports that switching threads is slow and that many threads keep the CPU busy (#11011 and others). OpenAI doesn't officially say that archiving makes things lighter, but archiving threads you've finished, with the CLI's /archive (which removes them from the list without deleting the records) or the app's archive feature, keeps the list from growing.
If you use scheduled tasks, the official troubleshooting guide includes steps for them. Frequent scheduled tasks pile up worktrees (copies of the working folder), so archive runs you no longer need, and don't pin runs you don't need to keep (OpenAI "Troubleshooting").
5. When running commands is slow (WSL and file location) command
Builds, tests, and searches that the AI requests run on your computer. If they're slow, check where your files are stored and your computer's performance, not the model or settings.
With WSL, keep files on the Linux side
If you run Codex in Linux inside Windows (WSL), the location of your working folder can make a big difference in speed. OpenAI explains that working under Windows drive paths, such as /mnt/c/..., can be slow, and recommends keeping repositories in your Linux home folder (for example, ~/code/my-app). If large repositories are slow, OpenAI also suggests increasing the memory and CPU allocated to WSL and updating WSL (OpenAI "WSL").
# Example %UserProfile%\.wslconfig: memory and CPU allocated to WSL
[wsl2]
memory=8GB
processors=4# Update WSL, then shut it down so the settings take effect (in PowerShell)
wsl --update
wsl --shutdownFor how to write .wslconfig, see Microsoft "Advanced settings configuration in WSL". wsl --shutdown stops all running WSL distributions, so close anything you're working on first. Set the values to match your computer's memory and CPU count.
Keep in mind that when you use an agent that runs natively on Windows in the Windows app, the opposite placement is recommended.
When the agent runs in WSL
Place files on the Linux side (such as ~/code)
Working under /mnt/c can be slow (OpenAI "WSL").
When the Windows app runs the agent natively on Windows
Place files in a Windows-side folder
The agent runs on the Windows side, so the slowness of /mnt/c doesn't apply. This is more reliable than opening a folder inside WSL directly from the app (OpenAI "ChatGPT desktop app for Windows"). Only when you work on the same repository with tools on the WSL side does access go through /mnt/<drive>/..., and those operations can be slow.
Source: OpenAI "ChatGPT desktop app for Windows". The Windows app defaults to an agent that runs natively on Windows. Switching to WSL requires restarting the app. In a report that the WSL agent was extremely slow (#25715), a staff member said a fix expected to ease the problem was merged on June 5, 2026, but as of October 25 the issue is still open.
Where computer performance matters
How long commands take depends on your computer's CPU, memory, and storage (especially whether it's an HDD or SSD). The movement of the app screen is also affected by your computer. The time the AI spends thinking and writing a reply, however, doesn't change when you buy a new computer. For what depends on your computer and how to measure where the time goes, see PC specs needed for Claude Code and Codex.
6. Reporting to OpenAI when nothing helps
If the problem persists after the steps above, you can report it to OpenAI, and staff can check the records to look for the cause.
- Send from the CLI: With
/feedbackin the CLI, you can send records and diagnostic information to the Codex team (follow the prompts and choose what to include). - Search GitHub: Search openai/codex issues for reports of the same symptom and staff replies. When you report, include your version number and OS.
- Capture detailed CLI logs: Plain-text logs are not created by default. If you start the CLI with a log location specified, such as
RUST_LOG=debug codex -c log_dir=./.codex-log(in Windows PowerShell,$env:RUST_LOG="debug"; codex -c log_dir=./.codex-log), acodex-tui.logfile is created (OpenAI "Environment variables").
Logs may contain the content of your work and file paths. Before you post them somewhere public, remove secret values and personal information.
Summary
- First check whether Codex is waiting for approval, what it's doing at the moment, and whether there's an outage.
- If replies are slow, lower the reasoning effort to match the task, switch to a lighter model, or use Fast mode (usage consumption goes up).
- Create a new thread for each task (as OpenAI staff recommend), and use
/compactif you need to continue. - If startup is slow, review your required MCP servers, stop unused servers with
enabled = false, and update the CLI. If that doesn't help, temporarily rename~/.codexto test. - If the app is heavy, update to the latest version and restart it, and archive scheduled task runs you no longer need.
- With WSL, keep files on the Linux side (or on the Windows side if you run natively on Windows).
- If the problem persists, send records with
/feedback, and search GitHub issues for the same symptom.
FAQ
Q. Is Codex slower than Claude Code?
There's no simple answer. Speed varies a lot depending on reasoning effort, the model, the type of task, app bugs, and how long local commands take. Few measurements compare the two under the same conditions, and the models are different, so it's risky to decide based on a single measurement. For a side-by-side explanation of which parts depend on your computer and which don't, see PC specs needed for Claude Code and Codex.
Q. Does using Fast mode always make things faster?
It speeds up how fast the model generates text, but it doesn't change how long commands take or how long MCP servers take to start. During busy periods on the server, things can still feel slow even in Fast mode (Chapter 2). Plan usage also runs out sooner, so it's safest to use it only for work that needs to be done quickly.
Q. What if the VS Code extension doesn't respond or is heavy?
Update the extension and VS Code, then restart VS Code. The official docs don't have specific steps for a heavy extension, but thread length (Chapter 2) also affects the extension in the same way. If the extension doesn't respond on Windows, OpenAI's documentation advises installing the Microsoft C++ Build Tools and the Visual C++ Redistributable package, then fully restarting VS Code (OpenAI "Windows sandbox").
