2026.10.20
Review Codex UltraFast settings when usage seems high
A personal example of noticing UltraFast was enabled and reconsidering the setting for work without active AI pair programming.
When Codex usage seemed to be dropping quickly, I noticed I was running it with UltraFast. I thought normal or Fast might be enough when I was not actively pair programming with AI, and asked for UltraFast to be left off when creating another thread.
In a follow-up later that day, I reported that the setting had been removed this time. This is an example of checking the selected setting and whether the work needs that speed when usage becomes a concern.
This records my own judgment and actions on October 19, 2026. I did not measure usage, savings, or output quality after the change. Whether normal or Fast is sufficient depends on the task; this is not a measured comparison of performance or savings between modes.
https://x.com/suna_gaku/status/2111968041353597112
Addendum: X article “Won’t you become Pro $500 too?”
I wrote an X article on when to use Ultrafast and tips for keeping usage consumption down.
“Won’t you become Pro $500 too? Figure out when to use Ultrafast and when not to, and get started with Ultrafast!” (Japanese)
From the article’s closing section:
- Use Ultrafast for work where a person waits on the AI’s reply (bouncing ideas on an implementation plan, fixing UI or copy)
- Ultrafast helps for urgent implementation and PoCs where you work alongside the AI
- Browser operations and API wait times don’t get faster, so leave testing to models like GPT-6.1 Terra
- Don’t run sub-agents or automations on Ultrafast; keep Ultrafast to the one thread you’re working in