You do not need a high-performance PC to run Claude Code or Codex. In both cases, the model that does the thinking runs on servers, so the time the AI spends thinking and writing will not get shorter when you buy a new PC. What PC performance changes is the time spent on commands that run locally, such as the builds, tests, and searches the AI asks for, and whether you have enough memory (if you do not, everything slows down).
Work time splits into two parts
Your PC
Local time
- Running the commands the AI asks for (builds, tests, search, Docker)
- The tools and apps themselves (whether memory is enough)
Changes with PC performance
Servers (Anthropic, OpenAI)
Model time
- The AI thinking, writing replies and code
- Speeding this up: fast modes (pricing and usage limits)
Does not change when you replace your PC
This article explains how to tell these two kinds of time apart, based on the official information for both tools and on a record of work kept on the author's PC. Where the two tools differ, the colors of Claude Code and Codex are used to show the difference side by side.
1. Official system requirements
These are the conditions described in the official documentation (confirmed October 24, 2026). Neither company publishes a recommended amount of memory. The 4GB for Claude Code is only the minimum needed to run.
"Windows-native" means running without WSL (Windows Subsystem for Linux). A sandbox is a mechanism that isolates the commands the AI runs. Sources: Claude Code "Set up Claude Code" and "Sandboxing"; OpenAI "Codex CLI," "ChatGPT desktop app," "ChatGPT desktop app for Linux," and "Sandbox"; the Codex CLI distribution files are on GitHub Releases.
2. Where the time goes — what changes with your PC
The ratios vary a lot with the type of work and the PC, so the most reliable approach is to measure the breakdown of your own work (see Section 6 for how). For reference, here is an example from records kept on the author's PC for the two tools.
Do not compare the two numbers directly (the denominators and the units of aggregation differ). Conditions: sessions that started between September 20 and October 24, 2026. PC: Intel Core i9-10900K, 48GB of memory, Windows 11; the working folder was on an HDD, and the work involved many scripts reading and writing many files. Claude Code: mainly Opus 5.5 sessions (sessions under one minute excluded). Codex: completed tasks using GPT-6.1 Sol and GPT-6 Astra (responses to one request; tasks that were stopped partway excluded). Neither record shows whether time spent waiting for approval is included (if it is, the local share would appear larger).
In this example from an HDD-based PC, local time accounted for a share that cannot be ignored in either tool. On a PC with an SSD, that share can be smaller. Only this part is likely to shrink on a faster PC; model time does not change when you replace the PC. How to decide whether to replace your PC is summarized in Section 8.
3. Speed comes from settings and pricing, not from your PC
When you want to reduce model time, what helps is the server-side settings and the model you use. Here are the fast modes of the two tools side by side (confirmed October 24, 2026).
Claude Code fast mode
Up to 2.5x faster
- Cost
- 2x the standard price ($8 input and $40 output per million tokens with Opus 5.5)
- Models
- Opus 5.5, Opus 5, Opus 4.8
- Who can use it
- Pro, Max, Team, Enterprise (paid from usage credits, not from the usage limit), and Claude Console
- How to switch
/fast
Codex Fast mode
Multiplier not stated
- Cost
- Uses 2.5x the standard plan usage limit (purchased credits count at 2x)
- Models
- GPT-6.1 Sol, GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, etc.
- Who can use it
- When signed in with ChatGPT
- How to switch
/fast(also in the desktop app and editor extensions)
Codex Ultrafast
Up to 8x faster (GPT-6 Astra)
- Cost
- Uses 8x the usage limit (purchased credits count at 6x)
- Models
- GPT-6 Astra, GPT-6.1 Sol (multiplier for Sol not stated)
- Who can use it
- Pro $500 (monthly) and eligible Enterprise and Edu plans
- How to switch
- Not stated (on eligible plans, check in model selection)
Sources: Claude Code "Speed up responses with fast mode," Anthropic "Pricing," and OpenAI "Speed."
There are also ways to reduce model time without paying more, in both tools.
| Method | Claude Code | Codex |
|---|---|---|
| Use a lighter model for simple tasks | Switch to Sonnet or Haiku with /model | Switch to GPT-6 Luna, etc., with /model |
| Lower the thinking strength (effort) | /effort | Choose it together with the model in /model |
| Reduce how much it reads | Name the scope precisely, such as "this function in this folder" / When the conversation gets long, summarize it or start a new one | |
Lowering thinking strength makes responses faster and cheaper; raising it takes longer but makes the model think more deeply (Claude Code "Model configuration," OpenAI "Speed"). For difficult tasks, a higher setting can improve quality, so switch depending on the task. Usage limits and extra charges are explained in more detail in "Claude Code token-saving tips and extra charges when you hit the limit" and "ChatGPT Pro plan comparison (pricing and usage limits)."
4. Where PC specs and how you use it matter
Memory — determined by what runs at the same time
You run the AI tool alongside your browser, editor, Docker, and development servers. The amount you need is set by the total of everything running at the same time, and if it runs short, swapping to disk (swap) slows everything down. The tools themselves also matter: Claude Code shows a warning when a single session goes above 2.5GB (Codex does not publish a guideline).
CPU and storage — builds, tests, and search
Builds and tests the AI runs take about as long as running them by hand. How fast files are searched and read depends on whether you have an SSD or an HDD, and on the number of files. Adding large build output folders to .gitignore so they are excluded from search also helps (a method Claude Code's official docs recommend).
Operations — avoid running heavy processes at the same time
If Docker builds or virtual machines are running on the same PC, the AI's commands compete with them for CPU and memory and end up waiting. When you ask for a long build or test, stopping other heavy processes shortens local time.
5. Using it on Mac, Windows, and Linux
Both tools run on all three operating systems. What differs by OS is whether the sandbox, the mechanism that isolates and runs the AI's commands, is available (confirmed October 24, 2026).
| Where it runs | Claude Code | Codex |
|---|---|---|
| Mac | Available Uses the macOS mechanism (Seatbelt); no extra setup needed | Available macOS mechanism (Seatbelt) |
| Windows-native | Not available Commands use Bash if Git for Windows is installed, otherwise PowerShell | Available Codex's own sandbox (table below) |
| WSL 2 | Available Requires bubblewrap and socat | Available Same mechanism as Linux (bubblewrap and seccomp) |
| WSL 1 | Not available Claude Code itself runs | Not supported Codex itself has not been supported since 0.115 |
| Linux | Available Requires bubblewrap and socat | Available bubblewrap and seccomp |
Sources: Claude Code "Sandboxing"; OpenAI "Sandbox" and "Windows sandbox" (confirmed October 24, 2026).
Claude Code Setting up the sandbox on Linux and WSL 2
Install bubblewrap and socat. On Ubuntu or Debian, use sudo apt-get install bubblewrap socat; on Fedora, use sudo dnf install bubblewrap socat. You can check whether anything is missing on the "Dependencies" tab of /sandbox in Claude Code. On a Mac, the macOS mechanism (Seatbelt) is used, so no extra setup is needed.
Codex Sandbox method when using Windows-native
One of three methods is chosen, in order from the top, based on what is available. The newest, MXC (Microsoft Execution Containers), uses the process isolation built into Windows. The documentation does not say whether Linux or WSL 2 require additional installation.
| Method | Conditions | Position |
|---|---|---|
| MXC | Windows 11 24H2 (OS build 26100.9278 or later) or 25H2 (OS build 26200.9278 or later), with the MXC features that Microsoft is rolling out in stages installed | Recommended (no administrator approval or firewall configuration required) |
| elevated | Default setting that an administrator approves | First fallback when MXC is unavailable |
| unelevated | Allowed by organization policy | Last fallback (weaker network isolation) |
You can check your OS build under Settings > System > About, in "Windows specifications." To check whether MXC actually works, run the following command in PowerShell from your project folder, using the Codex CLI (0.162.0 or later). If MXC_OK and exit code 0 appear, it works.
codex -c windows.sandbox=mxc sandbox --include-managed-config --permission-profile :workspace -- cmd.exe /d /c echo MXC_OK
$LASTEXITCODEWhether to use Windows-native or WSL 2 is basically decided by which toolchain your project runs on. If you use build tools for Linux, WSL 2 is the natural choice; if you build Windows apps, Windows-native is. If you want to use the sandbox with Claude Code, you need WSL 2.
6. Telling where the slowness comes from
Before replacing your PC, check where the time is going. The steps are the same in both tools.
- Check on screen where it is slow
The AI's text is slow to appear Model time. Check for outages on the status pages (Claude, OpenAI), then switch to a lighter model or lower the thinking strength (Section 3).A command shown as "running" never finishes Local time. Run that command in your own terminal. If it is also slow by hand, the problem is the command or your PC. Asking the AI to narrow the scope also helps.Typing or screen operations themselves are heavy A problem with the tool itself. Try the official fixes in Section 7.
- If local time is long, check the effect of replacing your PCLook at the three points in Section 8.
Measuring the breakdown with numbers (reading the record files, for experienced users)
| What to measure | Claude Code | Codex |
|---|---|---|
| Model time | "Total duration (API)" in /usage | Estimated as the total minus command execution time |
| Estimated local time | In the records (JSONL under ~/.claude/projects), the totalToolDuration in the cost-state line at the end of the session (also includes web fetches and so on) | In the records (JSONL under ~/.codex/sessions), the difference between started_at_ms and completed_at_ms in item_completed lines whose type is CommandExecution |
| Overall | "Total duration (wall)" in /usage (includes time waiting for input) | duration_ms in the task_complete line |
These are the same values used for the aggregation in Section 2. The locations of the records are given in each company's official documentation, but how to read the items is the author's own reading of the records and is not officially explained. Codex's thinking time (type Reasoning) does not include time spent writing replies or code, so estimate total model time by subtraction. Neither record shows whether time spent waiting for approval is included.
7. What to do when it is heavy or freezes
If the tool itself is heavy or freezes, try each tool's official fixes before worrying about PC performance. First, update the tool to the latest version, and if a session is long, close it and reopen it. This is the common first step for both.
Claude Code
- If a memory warning appears (heap above 2.5GB in one session), restart and continue with
claude --continue - When the conversation is long, use
/compact; restart between large tasks - Use
claude --safe-modeto compare with plugins, MCP servers, and hooks removed - If it stops responding, press Ctrl+C; if that fails, close the terminal and use
claude --resume - If your home folder is on a network drive (such as NFS), put programs on a local disk
The official Troubleshooting page covers the CLI. For the desktop app, start with updating and restarting.
Codex
- If it appears stuck, first check whether it is waiting for approval
- Check whether a simple command such as
git statusruns in the terminal - If not, start a new chat with a request narrowed to something small
- If the terminal freezes, reopen the panel; if it persists, restart the app after the chat finishes
- If scheduled runs add up many worktrees (Git's working copies), archive the runs you no longer need
The official Troubleshooting page is for the desktop app. For the CLI, start with updating and beginning a new session. For sluggishness in the Windows app, an OpenAI representative explained on GitHub that one of the causes was fixed in the August 28, 2026 build (ChatGPT v26.825.32147) (openai/codex #20214).
8. Three things to check before buying a new PC
If none of the following three points applies to you, replacing your PC will barely change how long your work takes. For the second point: if your working folder is on an HDD, moving it to an SSD first is a cheaper option.
Is memory running short?
During your usual work (with the AI tool, browser, editor, and Docker open), check with the method in the table below. If it is pinned at the limit, adding memory will help. For a new PC, use the usage at this point + the amount you want to add + some headroom as a guide.
Are there commands that are slow even when you run them by hand?
Find long commands with the method in Section 6, and time them by running them by hand. If they are long, a PC with a faster CPU or storage will shorten them. If you have an HDD, moving to an SSD first is the cheapest option.
Are you increasing the number of things you want to run at once?
If you want more sessions, subagents, or Docker containers running at the same time, memory and the number of CPU cores matter.
| What to check | Windows | Mac | Linux |
|---|---|---|---|
| Memory | Task Manager, "Performance" → "Memory": is in-use memory pinned at the limit? | Activity Monitor, "Memory" tab: is "Memory Pressure" yellow or red? | free -h: is available memory left? |
| SSD or HDD | Task Manager, "Performance" → "Disk" | "System Information," "Storage," "Media Type" | lsblk -d -o name,rota (HDD if ROTA is 1) |
9. If you run models locally, it is a different story
Everything above assumes Claude Code or Codex, where the model runs on a server. If you run a model on your own PC, for example with Ollama, PC performance translates directly into speed. In particular, GPU memory (VRAM) and system memory determine the size of model you can run and how fast it responds. For PC selection, see "Specs needed for local LLMs," and for getting started, see "How to run a local LLM."
10. Summary
- You do not need a high-performance PC. The Claude Code CLI needs 4GB of memory or more, x64 or ARM64, and an internet connection. Codex does not publish memory or CPU figures. Neither needs a GPU.
- Model time does not change with your PC. To make it faster, use fast mode (which uses more of your pricing or usage limits), a lighter model, lower thinking strength, or reduce how much is read.
- What your PC changes is local time (builds, tests, search, Docker). In the example from the author's HDD-based PC, it was a share that could not be ignored in either tool. The share varies by work and by PC, so look at your own breakdown before deciding.
- The OS changes the sandbox. On Windows-native, Claude Code is not available (it is with WSL 2), and Codex is available (the recommended method requires a recent Windows 11 build).
- When things are heavy or freezing, the tools' official fixes come first. Replace your PC only after you have found that local time is long.
FAQ
Q. Can I use it on a Mac with little memory or a cheap laptop?
A. The Claude Code CLI runs if you meet its requirements (see the table in Section 1), such as 4GB or more of memory and x64 or ARM64. Codex does not publish requirement figures, and the Mac desktop app's download is labeled for Apple silicon (the CLI also has a build for Intel Macs). With both tools, model time is the same as on a high-end PC. What makes a difference is local commands and memory when many apps are running at once.
Q. Is 8GB of memory enough?
A. Neither tool publishes a recommended amount. Claude Code's minimum (4GB) is exceeded, but whether it is enough depends on the total of what runs at the same time. As a guide, use your current usage during work, plus the amount you want to add, plus some headroom. For how to measure it, see the first point in Section 8.
Q. Will a new PC make things faster?
A. Model time will not change. Only local time, such as builds, tests, and search, gets shorter. Which one is longer depends on the work, so first check with the breakdown in Section 6.
Q. Do Claude Code and Codex require different PCs?
A. Not by much. Both run the model on servers, and your PC runs the tool itself plus the commands the AI asks for. The difference shows up in the sandbox: on Windows-native, Codex works (the recommended method requires a recent Windows 11 build), while Claude Code does not (it works with WSL 2).
Official pages referenced (confirmed October 24, 2026): Claude Code "Set up Claude Code," "Sandboxing," "Troubleshooting," "Speed up responses with fast mode"; OpenAI "Codex CLI," "ChatGPT desktop app," "Sandbox," "Windows sandbox," "Speed," "Troubleshooting." Pricing and requirements may change, so check the official pages before you use them.
