Optimizing GPT-6 Astra Usage in Codex
I have the Plus plan and have noticed that even moderately complex tasks can burn through Astra’s usage allowance pretty quickly. After hitting my limit, I spent $100 on additional usage so I could keep working. That worked well enough that, for my needs, buying extra usage occasionally seems more cost-effective than upgrading to the Pro plan. Below are a few more tips to help keep your Astra usage in check.
These settings are intended to reduce unnecessary usage while still taking advantage of GPT-6 Astra's capabilities.
1. Use Astra Low as your default
For most coding work, start with:
Model: GPT-6 Astra
Reasoning: Low
Speed: Standard
Move to Medium for more complicated debugging, architecture, or planning.
Use High only when a problem genuinely needs deeper reasoning or Low/Medium has failed to solve it.
Higher reasoning levels can consume substantially more tokens, and higher reasoning does not necessarily produce a better result.
2. Turn off Fast mode
In the Codex desktop app:
-
Start or open a Codex conversation.
-
Open the Intelligence control near the prompt box.
-
Find Speed.
-
Select Standard, not Fast.
Depending on the Codex app version, Speed may also appear in Settings.
If your organization manages Codex settings, an administrator can set the default under:
Workspace Settings → Models → Work & Codex → Speed
Why?
Fast mode makes Astra respond faster, but it uses significantly more of your Codex allowance. For most development work, Standard is the better choice.
3. Start fresh conversations periodically
Avoid keeping one enormous Codex conversation running for an entire project.
A useful approach is:
One substantial feature/problem → one Codex thread
For example:
-
Build authentication → one thread
-
Fix dashboard layout → new thread
-
Add CSV export → new thread
Long conversations accumulate context that Codex has to continue processing and eventually compact.
4. Disable automatic review?
Codex can use a separate model to automatically review certain approval requests.
Changing:
approvals_reviewer = "user"
means those approval decisions are routed to you instead of automatically invoking a reviewer.
The tradeoff is that Codex may occasionally stop and ask you to approve an action yourself.
This is probably off by default, but it may be worth checking.
5. Use another AI to help you write your prompt so your directions for Astra are clear and efficient
Place your outline into another AI. Tell it you would like to optimize token usage.
Recommended everyday workflow
Routine coding:
Astra Low + Standard (not fast)
Moderately difficult problem:
Astra Medium + Standard
Very difficult debugging/architecture:
Astra High + Standard
Need the answer faster and are willing to spend considerably more allowance:
Fast mode temporarily
The goal is not to avoid Astra's power. It is to use the expensive settings only when they provide a meaningful benefit.