codex - 💡(How to fix) Fix Codex Desktop can drive 60-80% GPU usage on M4 Max during subagent-heavy sessions [2 comments, 2 participants]

What version of Codex is running?

Codex Desktop 0.122.0-alpha.1

Which model were you using?

gpt-5.4 with high reasoning

What platform is your computer?

macOS on a MacBook Pro with M4 Max and 64 GB memory

What issue are you seeing?

Codex is using an unexpectedly large amount of GPU during a real coding session. On this machine I saw sustained GPU usage around 60-80%, which feels far out of proportion to the work being done.

I do not have direct GPU telemetry from the session logs themselves, but I do have the exact Codex session/thread that triggered it and the local session metadata.

What steps can reproduce the bug?

The exact session that triggered this was:

• Session/thread ID: 019d97b1-82c6-7d01-9042-030f94e105bc
• Thread name: Plan audit fix proposal
• Session start: April 16, 2026 at 19:08:06 UTC
• Source: vscode
• Working directory at the time: /Users/treygoff/Code/atlasos

From the local session trace, this thread then spawned four subagents in quick succession at approximately April 16, 2026 19:09:28 UTC:

• Descartes (explorer)
• McClintock (explorer)
• Lagrange (explorer)
• Plan Critic (plan_checker)

All of those subagents were launched with gpt-5.4 and high reasoning.

A plausible repro pattern seems to be:

1. Use Codex Desktop on macOS in a VS Code-originated session.
2. Start a task that triggers multiple high-effort subagents in parallel.
3. Keep the session open while they run and report back.
4. Observe GPU usage in Activity Monitor.

On my machine, GPU usage climbed to roughly 60-80% during this workload.

What is the expected behavior?

Even when a session is doing substantial reasoning or coordinating several subagents, local GPU usage should stay much lower unless there is a clearly necessary rendering workload.

What do you see instead?

Codex appears to drive very high GPU usage during this session pattern, despite the task being primarily text/orchestration work.

Additional information

A few concrete data points from the session metadata:

• Parent thread model: gpt-5.4
• Parent thread reasoning: high
• Parent thread source: vscode
• The issue was observed on an M4 Max MacBook Pro with 64 GB RAM, so this is not extremely constrained hardware.

If useful, I can provide more local session artifacts tied to 019d97b1-82c6-7d01-9042-030f94e105bc.