SilverleafSolutions

AI Explorer. Engineering better software, asking deep questions. Founder, Silverleaf Solutions LLC

[SilverleafSolutions @SilverleafTech]

Codex usage limits are back to 5x/20x vs Plus. Here are some tips to stretch your quota:

-Always use plan mode before allowing Codex to do work
-Disable Codex memories in Personalization settings
-Turn off noisy build warnings
-Keep AGENTS.md small
-Disable unused MCPs
-Use smallest model/low reasoning for routine edits

[SilverleafSolutions @SilverleafTech]

AI is moving from "help me write code" to "help me turn an idea into something usable."

Codex Sites turns plans and rough concepts into interactive apps your team can actually explore, share, and refine. That points to a much faster loop between thinking, building, and testing.

[OpenAI @OpenAI]

Building apps has never been easier.

With Sites, Codex can turn your work, ideas, and plans into an interactive website or app your team can explore, use, and share with a URL.

Rolling out to Business and Enterprise plans, before expanding more broadly.

[SilverleafSolutions @SilverleafTech]

@tinkerersanky @OpenAI @thsottiaux Case in point: I'm on the $100/mo plan, and I had Codex spinning on tasks for hours yesterday, and I went from 86% to 72% weekly usage. So it will pay huge dividends to do research and figure out what's causing so much usage!

[SilverleafSolutions @SilverleafTech]

@tinkerersanky @OpenAI @thsottiaux This just simply isn't the case for everyone. You likely have something eating up your quota without you knowing it. See if any of these help: https://t.co/Skr3lVFYlD

[SilverleafSolutions @SilverleafTech]

CODEX cheatsheet to reduce quota usage:

1. Scope the task tightly.
Tell Codex exactly what to change, what not to touch, and how to verify it. Vague tasks burn tokens.

2. Use plan mode for unclear or multi-file work.
Skip it for tiny edits. Use it when architecture, scope, or file impact is uncertain. Helps reduce quota usage by not burning expensive output tokens on unnecessary edits.

3. Keep AGENTS.md short and precise.
Give Codex the repo map, commands, conventions, and hard rules. If it gets long, keep the main file minimal and link to task-specific docs.

4. Clean up dead code and old projects.
Deprecated folders, abandoned prototypes, and unused projects create false trails. Codex may inspect them unless your repo clearly says they are irrelevant.

5. Ignore the right things.
Use .gitignore/ignore rules for bin, obj, node_modules, generated clients, coverage output, logs, minified bundles, etc.

6. Disable unused MCPs.
Every extra tool can add context and decision overhead. Keep only the tools that actively help your workflow.

7. Reduce noisy build/test output.
Fix real warnings. Suppress or isolate known irrelevant noise so Codex does not chase unrelated issues.

8. Use the smallest model/reasoning that fits.
Low/medium for routine edits. High/xhigh for hard debugging, refactors, architecture, or messy unknowns.

9. Be careful with saved context
If Codex memories or long-running thread context are adding stale/cross-project noise, start a fresh task or turn them off. Put durable repo rules in AGENTS.md instead.

10. Prefer clean signal over less context.
The goal is not starving Codex to save tokens. The goal is giving it the fewest high-quality clues needed to do the job right.

What would you add? Reply with anything I missed.

[SilverleafSolutions @SilverleafTech]

@glvnme Great thoughts, thanks for sharing! 🦾

[SilverleafSolutions @SilverleafTech]

Codex usage limits are back to 5x/20x vs Plus. Here are some tips to stretch your quota:

-Always use plan mode before allowing Codex to do work
-Disable Codex memories in Personalization settings
-Turn off noisy build warnings
-Keep AGENTS.md small
-Disable unused MCPs
-Use smallest model/low reasoning for routine edits

[SilverleafSolutions @SilverleafTech]

@CryptoValueLabs Yeah, you are probably right. At this point, I think they want to get everyone hooked on Codex, and they've largely accomplished that at this point.

[SilverleafSolutions @SilverleafTech]

@D444433322 Yes! What kind of things did it do after you told it to do this?

[SilverleafSolutions @SilverleafTech]

@CryptoValueLabs Love this! Feel free to share the additional ones!

[SilverleafSolutions @SilverleafTech]

CODEX cheatsheet to reduce quota usage:

1. Scope the task tightly.
Tell Codex exactly what to change, what not to touch, and how to verify it. Vague tasks burn tokens.

2. Use plan mode for unclear or multi-file work.
Skip it for tiny edits. Use it when architecture, scope, or file impact is uncertain. Helps reduce quota usage by not burning expensive output tokens on unnecessary edits.

3. Keep AGENTS.md short and precise.
Give Codex the repo map, commands, conventions, and hard rules. If it gets long, keep the main file minimal and link to task-specific docs.

4. Clean up dead code and old projects.
Deprecated folders, abandoned prototypes, and unused projects create false trails. Codex may inspect them unless your repo clearly says they are irrelevant.

5. Ignore the right things.
Use .gitignore/ignore rules for bin, obj, node_modules, generated clients, coverage output, logs, minified bundles, etc.

6. Disable unused MCPs.
Every extra tool can add context and decision overhead. Keep only the tools that actively help your workflow.

7. Reduce noisy build/test output.
Fix real warnings. Suppress or isolate known irrelevant noise so Codex does not chase unrelated issues.

8. Use the smallest model/reasoning that fits.
Low/medium for routine edits. High/xhigh for hard debugging, refactors, architecture, or messy unknowns.

9. Be careful with saved context
If Codex memories or long-running thread context are adding stale/cross-project noise, start a fresh task or turn them off. Put durable repo rules in AGENTS.md instead.

10. Prefer clean signal over less context.
The goal is not starving Codex to save tokens. The goal is giving it the fewest high-quality clues needed to do the job right.

What would you add? Reply with anything I missed.

[SilverleafSolutions @SilverleafTech]

@dkhang My understanding is that Claude has had worse quota problems than Codex. 😬

[SilverleafSolutions @SilverleafTech]

@glvnme Nice! I'll try that a bit more. Another thing I've found super useful: ask Codex to interview you about the task. The questions it asks are typically awesome, revealing extra details you may have missed. 👍

[SilverleafSolutions @SilverleafTech]

Interesting. The plans are incredibly useful, in my experience, as I can see where the LLM wants to head and what it wants to edit. I've found Codex to be far-reaching in what it detects it needs to change and nearly always close to what I would have wanted done. But plan mode gives me the change to redirect it and fine-tune the goal and remind it of other priorities or areas of the system it needs to address or consider. 👍

[SilverleafSolutions @SilverleafTech]

@glvnme Interesting thought about token efficiency. 5.5 is 2x more expensive than 5.4. I wonder if it's basically even after considering 5.5 efficiency gains? Not sure.

[SilverleafSolutions @SilverleafTech]

@glvnme Plan mode is a massive help to better focus Codex and make it explain to you what it wants to do before it starts using output tokens to do work. Output tokens are 6x more expensive than input! 😬

[SilverleafSolutions @SilverleafTech]

@erikhansen 🤣

[SilverleafSolutions @SilverleafTech]

@erikhansen Yes! Without verification, one could be throwing tokens in the trashcan. Ain't nobody got time for that.

[SilverleafSolutions @SilverleafTech]

We're all frustrated that Codex retired the 2x quota limits today. But use this as an opportunity to become a master of Codex! Figure out how it works. Understand how your prompts change its behavior. Ask it how to reduce token usage too! And don't forget to have fun. 🙂

[SilverleafSolutions @SilverleafTech]

Since many seem to find this post helpful, I wrote an even longer cheatsheet with the top 10 ways to reduce Codex quota usage. https://t.co/Skr3lVFYlD

[SilverleafSolutions @SilverleafTech]

CODEX cheatsheet to reduce quota usage:

1. Scope the task tightly.
Tell Codex exactly what to change, what not to touch, and how to verify it. Vague tasks burn tokens.

2. Use plan mode for unclear or multi-file work.
Skip it for tiny edits. Use it when architecture, scope, or file impact is uncertain. Helps reduce quota usage by not burning expensive output tokens on unnecessary edits.

3. Keep AGENTS.md short and precise.
Give Codex the repo map, commands, conventions, and hard rules. If it gets long, keep the main file minimal and link to task-specific docs.

4. Clean up dead code and old projects.
Deprecated folders, abandoned prototypes, and unused projects create false trails. Codex may inspect them unless your repo clearly says they are irrelevant.

5. Ignore the right things.
Use .gitignore/ignore rules for bin, obj, node_modules, generated clients, coverage output, logs, minified bundles, etc.

6. Disable unused MCPs.
Every extra tool can add context and decision overhead. Keep only the tools that actively help your workflow.

7. Reduce noisy build/test output.
Fix real warnings. Suppress or isolate known irrelevant noise so Codex does not chase unrelated issues.

8. Use the smallest model/reasoning that fits.
Low/medium for routine edits. High/xhigh for hard debugging, refactors, architecture, or messy unknowns.

9. Be careful with saved context
If Codex memories or long-running thread context are adding stale/cross-project noise, start a fresh task or turn them off. Put durable repo rules in AGENTS.md instead.

10. Prefer clean signal over less context.
The goal is not starving Codex to save tokens. The goal is giving it the fewest high-quality clues needed to do the job right.

What would you add? Reply with anything I missed.

[SilverleafSolutions @SilverleafTech]

Sure!
-Keep the repo clean by removing old/deprecated code or projects.
-Keep AGENTS.md short and precise, so Codex can navigate your repo more efficiently. If it gets too lengthy, minimize the main file and link to task-specific docs.
-Use .gitignore/ignore rules so Codex does not waste time on bin, obj, node_modules, generated clients, coverage output, logs, minified bundles, etc.

[SilverleafSolutions @SilverleafTech]

Some are correctly noting that saying "always use plan mode" is overkill. A more accurate statement would be "always use plan mode when the task is complex, ambiguous, or hard to describe.

[SilverleafSolutions @SilverleafTech]

@yyhh6tgg63536 I read that recently and was shocked because I change reasoning level all the time mid-thread and simply have not dealt with large-quota usage! I'm genuinely confused why Codex uses so much less quota for me. 🤷‍♂️