- My Forums
- Tiger Rant
- LSU Recruiting
- SEC Rant
- Saints Talk
- Pelicans Talk
- More Sports Board
- Fantasy Sports
- Golf Board
- Soccer Board
- O-T Lounge
- Tech Board
- Home/Garden Board
- Outdoor Board
- Health/Fitness Board
- Movie/TV Board
- Book Board
- Music Board
- Political Talk
- Money Talk
- Fark Board
- Gaming Board
- Travel Board
- Food/Drink Board
- Ticket Exchange
- TD Help Board
Customize My Forums- View All Forums
- Show Left Links
- Topic Sort Options
- Trending Topics
- Recent Topics
- Active Topics
Started By
Message
UPDATED - my codex token saving plan and its working!
Posted on 8/27/26 at 5:36 am
Posted on 8/27/26 at 5:36 am
This is specific to chatgpt/codex but this combo has dropped my token use DRAMATICALLY. Like, night and day.
for my 'television network' my dumb arse was asking codex to build each tv show and keep git updated.
i didn't realize 'change that shade of blue on the background' meant that codex was taking the entire car apart piece by piece then reassembling it and then finishing it off with that nice little swipe of blue paint. i was burning my weekly allowance with one dashboard/tv show edit request.
i have about 20 'tv shows' queued up so this was clearly unsustainable even with the $100/mo plan which only gives you about 5x as many tokens..so great....now i run out in 5 hours instead of 1.
*****************************
2 changes have dramatically improved my token consumption issue.
1. Hybrid gdrive/local git approach - chatgpt 'normal' chat can access and edit gdrive files all day and night 24/7 and seem to burn little to no tokens. this is my 'hybrid' approach now so i can get my coding requirements in order well ahead of time and codex knows to consult both before suggesting any updates.
2. with all the shared templates and context, codex will just do one targeted code update and the scale is much smaller. No more 151 files getting updated and git revisions to change the background color.
Prototype dash/tv show, including a polished whole-house battery-health scene.
Design the persistent security ribbon/chiron behavior visually.
Batch all dash visual feedback rather than feeding Codex one tweak at a time.
I'll have complete, approved show briefs and mockups sitting in Drive waiting for implementation.
Now when Codex reopens, instead of: “Make me a Good Morning dashboard.”
it receives:
“Class C ticket. Here is the approved Good Morning brief and mockup. Consume these three known HHTV interfaces. Create these show-owned files. Shared shell/controller are read-only. Plan the file set first.”
for my 'television network' my dumb arse was asking codex to build each tv show and keep git updated.
i didn't realize 'change that shade of blue on the background' meant that codex was taking the entire car apart piece by piece then reassembling it and then finishing it off with that nice little swipe of blue paint. i was burning my weekly allowance with one dashboard/tv show edit request.
i have about 20 'tv shows' queued up so this was clearly unsustainable even with the $100/mo plan which only gives you about 5x as many tokens..so great....now i run out in 5 hours instead of 1.
*****************************
2 changes have dramatically improved my token consumption issue.
1. Hybrid gdrive/local git approach - chatgpt 'normal' chat can access and edit gdrive files all day and night 24/7 and seem to burn little to no tokens. this is my 'hybrid' approach now so i can get my coding requirements in order well ahead of time and codex knows to consult both before suggesting any updates.
2. with all the shared templates and context, codex will just do one targeted code update and the scale is much smaller. No more 151 files getting updated and git revisions to change the background color.
Prototype dash/tv show, including a polished whole-house battery-health scene.
Design the persistent security ribbon/chiron behavior visually.
Batch all dash visual feedback rather than feeding Codex one tweak at a time.
I'll have complete, approved show briefs and mockups sitting in Drive waiting for implementation.
Now when Codex reopens, instead of: “Make me a Good Morning dashboard.”
it receives:
“Class C ticket. Here is the approved Good Morning brief and mockup. Consume these three known HHTV interfaces. Create these show-owned files. Shared shell/controller are read-only. Plan the file set first.”
quote:
Build Weather on the Hour from the approved brief and mockups.
Read the Weather brief, Data Contracts weather rows, and the documented HHTV shell interface only.
Meteocons is existing shared read-only infrastructure.
Do not scan the repo.
Do not modify the shell/controller/global CSS.
First report the proposed Weather-owned file set and direct dependencies. Do not code yet.
This post was edited on 9/3/26 at 4:11 pm
8
Posted on 8/27/26 at 7:44 am to CAD703X
I hit my limit with Claude in about 45 minutes, doing stuff I could spend all day in ChatGPT doing with no warnings. So, I told Claude I needed a prompt to have ChatGPT work like Claude was doing for that task. I used the prompt on ChatGPT and it started working just like Claude (not coding based). No more token problems.
Posted on 8/27/26 at 8:33 am to CAD703X
Lower the model level according to the task. I'm using haiku 4.5 for all my smart home/home assistant stuff and it's pennies each time. Also, I use the free version of claude for basic shite and reserve the claude platform for only code-based chats.
Posted on 8/27/26 at 8:44 am to CAD703X
Saving tokens is impossible outside of turning thinking down. Saving your context window is all the rage. For that there is graphify and sub agents.
Personal accounts simply don’t give you enough tokens to get things done, even at $200 per month, unless you schedule your life around the usage windows.
If you want to get things done on your own schedule, you need to move over to openrouter and use mainly deepseek 4 flash or Luna. You can still click over to something more expensive if you get in a bind and never have to leave your coding agent.
Edit: obv context and tokens closely related but there is still no one-size-fits-all trick. Don’t pull datasets directly on small projects. Don’t go into huge repos without graphs. As far as the output tokens, that’s 90% thinking. Maybe more. Don’t allow big think when it doesn’t need it.
Personal accounts simply don’t give you enough tokens to get things done, even at $200 per month, unless you schedule your life around the usage windows.
If you want to get things done on your own schedule, you need to move over to openrouter and use mainly deepseek 4 flash or Luna. You can still click over to something more expensive if you get in a bind and never have to leave your coding agent.
Edit: obv context and tokens closely related but there is still no one-size-fits-all trick. Don’t pull datasets directly on small projects. Don’t go into huge repos without graphs. As far as the output tokens, that’s 90% thinking. Maybe more. Don’t allow big think when it doesn’t need it.
This post was edited on 8/28/26 at 4:05 pm
Posted on 8/27/26 at 10:40 am to CAD703X
I have become very cynical about how LLM's waste tokens by not following instructions, hallucinating, or not doing what you asked of them.
I absolutely have come to believe this is by design to drive up token usage.
This is a product that I pay for, when it doesn't follow instructions and creates mistakes that require further token usage to fix, why should I bear that financial cost.
ETA at work we have Claude Enterprise with a $10K monthly usage limit, I haven't hit that yet, but I routinely go over $2k a month.
I absolutely have come to believe this is by design to drive up token usage.
This is a product that I pay for, when it doesn't follow instructions and creates mistakes that require further token usage to fix, why should I bear that financial cost.
ETA at work we have Claude Enterprise with a $10K monthly usage limit, I haven't hit that yet, but I routinely go over $2k a month.
This post was edited on 8/27/26 at 10:43 am
Posted on 8/27/26 at 10:41 am to CAD703X
Sorry you’ll have to start thinking for yourself again baw
Posted on 8/27/26 at 11:34 am to Chromdome35
quote:i went down a rabbit hole yesterday when my supposedly 'hardened' debian chromium display PC browser crashed in the middle of the night.
I have become very cynical about how LLM's waste tokens by not following instructions, hallucinating, or not doing what you asked of them.
I absolutely have come to believe this is by design to drive up token usage.
rather than chalk it up to a fluke or simply set a task to reboot before the display comes online every morning at 6am, i had codex ssh into the miniPC and spent 2 hours having it run diagnostics, download and scan logs...essentially i let it use all my tokens chasing some invisible issue that was likely no more than a fluke in the first place
AI will happily waste your time and money if you let it.
Posted on 8/27/26 at 11:53 am to CAD703X
quote:
AI will happily waste your time and money if you let it.
It did what you told it to do
Posted on 8/27/26 at 12:44 pm to Mingo Was His NameO
quote:y r u the way u r
what you told it to do
Posted on 8/27/26 at 2:47 pm to LemmyLives
quote:
I hit my limit with Claude in about 45 minutes, doing stuff I could spend all day in ChatGPT doing with no warnings.
I hit my limit today in 15 mins doing the same stuff that i've been doing all week that would last the whole window. I think they fricked with something.
Posted on 8/27/26 at 7:06 pm to CAD703X
I assume you're using standard token minimization strategies such as limiting conversation length, not loading up big data sets, etc...
Posted on 8/27/26 at 10:37 pm to CAD703X
There’s some skills you can load that will help with that. Before I started customizing my Claude, I would run into similar issues. Also, I’ll try and find the prompt I made to cut all that shite out.
Now I’m coding with Claude and I’ve built 3 apps and I can code for a couple hours before I hit a usage limit. I’ll even code and run cowork stuff for my classes or projects I’m working on.
Next thing I’m going to try is connecting Claude with arcGIS.
Now I’m coding with Claude and I’ve built 3 apps and I can code for a couple hours before I hit a usage limit. I’ll even code and run cowork stuff for my classes or projects I’m working on.
Next thing I’m going to try is connecting Claude with arcGIS.
Posted on 8/28/26 at 3:19 pm to Chromdome35
I found my problem. I forgot i switched back down to the pro plan last month. I was paying for less usage. Back to Max.
Posted on 9/3/26 at 1:10 pm to CAD703X
bump updated OP.
tl;dr gdrive access is a Godsend; chatgpt can read/write/edit there until the cows come home w/ virtually zero token hit (for now) and codex knows this is part of the project it has to consult when coding.
its like a cheat code for codex to have chatgpt populate all the docs, spreadsheets, etc online for free so codex doesn't have to.
tl;dr gdrive access is a Godsend; chatgpt can read/write/edit there until the cows come home w/ virtually zero token hit (for now) and codex knows this is part of the project it has to consult when coding.
its like a cheat code for codex to have chatgpt populate all the docs, spreadsheets, etc online for free so codex doesn't have to.
Posted on 9/4/26 at 7:33 pm to CAD703X
quote:
tl;dr gdrive access is a Godsend; chatgpt can read/write/edit there until the cows come home w/ virtually zero token hit
What makes that different compared to working off of your PC?
Posted on 9/4/26 at 8:27 pm to TigerFanatic99
quote:i burn tokens writing files to the pc because i'm forced to switch to 'work' mode to access local resources.
What makes that different compared to working off of your PC?
i burn no tokens writing to gdrive at least so far i have done an insane amount of storing images, spreadsheets, etc. and it keeps everything in sync and never have to switch out of chat mode.
i tried to have it write a word doc to my pc and it said 'no tokens until next week' so i said fine, lets just keep writing in the cloud.
codex knows to use those docs on gdrive as part of its repo memory and thats where my project lives so its fantastic its keeping all those documents in sync. i don't have to worry about a thing or token use. its not slowing me down until i need to execute some code and i have plenty to work on before then.
eta bonus: if needed i could access them on my pc by installing google drive on it and 'cheat' that way as well but haven't found i needed to do that yet.
eta this is a huge loophole and i bet chatgpt comes along and shuts it down at some point.
This post was edited on 9/4/26 at 8:32 pm
Page 1 of 1
Popular
Back to top
Follow TigerDroppings for LSU Football News