ChatGPT macOS app: ~4 GB idle `codex` process makes 8 GB Mac unusable

Summary

The current ChatGPT macOS app starts a codex process that consumes roughly 3.6–3.9 GB of memory immediately after launch, even when no Work/Codex conversation or project is opened. On an 8 GB Apple Silicon Mac mini this causes heavy memory compression, swap usage, red memory pressure, and severe system/app slowdown.

This was reproduced repeatedly in Activity Monitor while comparing the current ChatGPT app against ChatGPT Classic on the same machine.

Environment

• Device: Apple Silicon Mac mini
• Physical memory: 8 GB
• Current ChatGPT app: approx. 1.28 GB application size
• Reported app/version metadata: Codex 151.0.7922.76
• ChatGPT Classic comparison build: 1.2026.184
• Date observed: 2026-08-13
• Subscription/use case: ChatGPT Business; Work/Codex/browser functionality is important for working with authenticated web tools such as n8n

Reproduction

1. Quit the current ChatGPT app completely (Cmd+Q).
2. Wait until its codex, Codex Renderer, and Codex Service processes disappear.
3. Observe Activity Monitor: memory pressure returns to green.
4. Launch the current ChatGPT app.
5. Do not open an existing Work/Codex chat or project; leave the app essentially idle/new-chat state.
6. Observe Activity Monitor.

Observed results

Repeated measurements included:

• codex: ~3.59 GB immediately after launch
• codex: ~3.77 GB after launch
• codex: ~3.78 GB after several minutes idle (so it stabilizes rather than continuously leaking)
• codex: ~3.94 GB in a clean test with no existing Work chat opened
• ChatGPT process itself: roughly 300–950 MB depending on the run
• Additional Codex Renderer/Service processes: several hundred MB more
• Total memory in one clean run: ~6.63 GB / 8 GB
• Compressed memory: ~2.65 GB
• Swap: ~1.87 GB shortly after launch
• In another run after use, swap reached ~9.26 GB and memory pressure was red

Quitting only the current ChatGPT app caused an immediate step down in memory pressure. One post-quit measurement showed:

• Memory used: ~3.87 GB
• Compressed: ~756 MB
• Swap: ~504 MB
• Memory pressure: green
• all Codex processes gone

Control test: ChatGPT Classic

On the same Mac, ChatGPT Classic (1.2026.184) was launched instead.

Observed:

• ChatGPT: ~117 MB
• ChatGPTHelper: ~26 MB
• no multi-GB codex process
• compressed memory: ~656 MB
• swap: ~328 MB
• memory pressure remained green

This makes the difference between the current Codex-integrated app and Classic very clear on an 8 GB machine.

Expected behavior

Launching ChatGPT while idle should not consume approximately half of an 8 GB Mac's physical memory in the codex backend before the user starts a Work/Codex task. The app should remain usable on supported 8 GB Apple Silicon Macs, or Codex/Work components should be lazy-loaded/on-demand so normal chat use does not incur this memory cost.

Actual impact

The current app becomes impractical for normal work on this 8 GB Mac because macOS rapidly compresses memory and swaps to disk. The web app and ChatGPT Classic remain responsive, but they are not equivalent replacements for the user's workflow because the current desktop app's Work/Codex/browser capabilities are needed for working directly with authenticated tools such as n8n.

The user also has local Work conversations that are not visible in the web chat history, making simply abandoning/resetting the desktop app undesirable.

Request

Please investigate the baseline memory footprint of the codex process in the current ChatGPT macOS app, especially on 8 GB Apple Silicon systems. In particular, please consider lazy-loading or suspending Codex/Work services until they are actually needed.

This is a reproducible usability blocker rather than gradual memory growth: the ~3.6–3.9 GB codex allocation appears immediately on launch and remains around that level while idle.

No source file, test, or entry point is named in the report. First reproduce the idle launch measurements on an 8 GB Apple Silicon Mac and trace which Codex/Work services start with the macOS app; done means identifying the baseline allocation and establishing that idle launch no longer causes the reported multi-GB footprint or severe memory pressure.