19:19Dubibubi · Review
I Made Claude Opus 5 and Fable 5 Build the Same App
A blind, three-build test of Anthropic's newest coding models — and the first time this reviewer picked a model other than Fable as the winner.
July 25thThe argument in one line.
GPT-6 Astra changed how Codex spends tokens, so the old context-limit habits are dead, and the real savings now come from usage budgets, cheap sub-agents, and shorter reports.
Read if. Skip if.
READ IF YOU ARE…
- You use OpenAI Codex or a similar coding agent daily and keep hitting weekly or five-hour usage limits.
- You manage AI coding costs for yourself or a small team and want concrete percentage numbers, not vague advice.
- You already use AGENTS.md files, skills, or sub-agents and want to tune them rather than learn them from zero.
SKIP IF…
- You use Claude Code, Cursor, or another tool with a different pricing and context model. These rules are specific to Codex and GPT-6 Astra.
- You've never used an AI coding agent before. This assumes you already know what a token, context window, and tool call are.
The full version, fast.
GPT-6 Astra rewired how OpenAI's Codex spends tokens, so the token-saving habits built for older models, like keeping context under 270,000 tokens or watching five-minute cache resets, no longer do anything. The real savings now come from treating usage like a budget: giving Astra a percentage target and letting it run experiments to hit it, delegating routine steps to the far cheaper Luna Max sub-agent instead of running everything through Astra directly, saving preferences once in AGENTS.md instead of re-explaining them every session, and keeping tool-call reports and final answers short by default. Measured tests in the video show giving Astra the exact file path cut tokens by about 23%, and a full vanilla-versus-optimized run on the same bug came in 33.3% cheaper.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:42
01 · The 91.75% claim
The creator says he used to hit Codex usage limits daily, and after changing how he uses it, hasn't hit a limit in over a week.
00:42 – 01:45
02 · Why the old token tips are dead
A quiz on popular pre-Astra token-saving habits, capping context at 270,000 tokens, watching 5-minute cache resets, reveals none of them still apply now that the cache window is 30 minutes and effort-level changes no longer force a re-read of the chat.
01:45 – 03:36
03 · Rule 1: give Astra a usage budget
Because Astra can see its own usage limits, it can be given an explicit percentage target to stay within, and one workflow went from 25 products per allowance to 300 once it was told to optimize against a 1% budget.
03:36 – 04:57
04 · Rule 2: delegate to Luna Max sub-agents
Astra can orchestrate the far cheaper Luna Max model for narrow, well-defined steps and only step in itself to review the output or handle anything complicated.
04:57 – 07:41
05 · Rules 3-5: cron trick, reusable skills, saved preferences
A 6am cron ping starts the 5-hour usage window before the workday begins, repeatable processes get turned into reusable skills instead of being re-explained every time, and preferences get saved once, personal ones in account settings, project ones in AGENTS.md.
07:41 – 09:42
06 · Rules 6-7: steer mid-run, turn off unused tools
Switching follow-up behavior to steer lets a correction reach a run while it's still working, and every switched-on plugin, MCP server, or image-generation call costs usage whether it's used in that session or not.
09:42 – 12:04
07 · Rules 8-9: shorter reports, shorter answers
Trimming tool-call reports down to pass or fail cut tokens about 6% in testing, and asking for short answers by default instead of long explanations cuts the output side of the bill.
12:04 – 13:18
08 · Rules 10-11: exact file paths, progress records
Giving Codex the exact file path for a bug instead of making it search cut tokens about 23%, and keeping a short running progress file lets a task resume cleanly after an interruption or crash.
13:18 – 16:20
09 · Live demo: vanilla versus enhanced Codex
The same button-fixing bug is run twice, once with a completely default Codex prompt and once with a budget, sub-agent delegation, trimmed plugins, and an exact file path, landing a 33.3% token saving.
Lines worth screenshotting.
- None of the popular pre-Astra token-saving tricks, like capping context at 270,000 tokens, still do anything on GPT-6 Astra.
- Astra's cache window now lasts 30 minutes instead of 5, and changing effort levels mid-conversation no longer forces a full re-read of the chat.
- Because Astra can see its own usage limits, you can tell it to finish a task within 3% of your weekly allowance and it will work to hit that target.
- One workflow went from using 4% of a weekly usage allowance to getting three results done for just 1%, once it was given an explicit budget to optimize against.
- Delegating routine steps to the cheaper Luna Max sub-agent instead of running them through Astra directly can cut those steps' cost by roughly 98%.
- Luna Max is benchmarked as equivalent to Sonnet 5 at max for about a seventh of the price, after a reported 80% price cut.
- AGENTS.md is read at the start of every task, so instructions saved there stop you from re-teaching a coding agent the same preferences in every new chat.
- A saved instruction file still costs context on every single task, so bulky, unused instructions quietly burn tokens even when they're never the actual problem.
- At the start of a task, Codex only reads a skill's name and one-line description, and only loads the full instructions once it decides to use it, so unused skills cost almost nothing to keep around.
- Switching a mid-run correction from 'queue for next run' to 'steer the current run' stops an agent from finishing an entire response you're about to throw away.
- Every tool switched on, plugins, image generation, retrieval, caching, counts against usage whether it gets used in that session or not.
- Image generation draws from the same usage pool as chat but burns through it three to five times faster per request.
- Trimming tool-call reports down to pass or fail instead of full write-ups cut tokens by about 6% in matched comparisons, with no missed problems.
- Giving a coding agent the exact file path where a bug lives, instead of asking it to search the whole codebase, cut total tokens by about 23% in testing.
- A live side-by-side test fixing the same button bug showed an optimized Codex configuration using 33.3% fewer tokens than a completely default run.
How AI agent usage limits actually work
USAGE ECONOMICS
Usage limits reward specificity and delegation: telling the model what it's allowed to spend, who should do the cheap work, and how much detail you actually need back.
02Why the old token tips are dead
- Token-saving habits tuned for one model generation can become useless overnight when the underlying model changes, so check whether an old habit still applies before trusting it out of routine.
- A cache window moving from 5 minutes to 30 minutes changes how much you need to worry about a short gap between sessions costing you a full reprocessing charge.
03Rule 1: give Astra a usage budget
- If a coding agent can see its own usage limits, you can give it an explicit percentage budget for a task instead of hoping it works efficiently by default.
- Simply mentioning a token or usage budget in a prompt tends to make a model behave more conservatively, even before you ask it to actively optimize anything.
03Rule 1: give Astra a usage budget (case study)
- Letting an agent run its own experiments aimed specifically at hitting a usage target, rather than just completing the task, can produce order-of-magnitude efficiency gains on repeatable work.
- Small per-run savings compound fast at scale: a large efficiency gain on one workflow step multiplies across every future run of that same workflow.
04Rule 2: delegate to Luna Max sub-agents
- Not every step of a plan needs your most expensive model to execute it personally; a cheaper sub-agent can handle narrow, well-scoped steps for a fraction of the cost.
- Give a cheaper sub-agent a narrow job and have the expensive model review the output and handle only the complicated parts at the end.
05Rules 3-5: cron trick, reusable skills, saved preferences
- If your usage window resets on a fixed timer, pinging it before your actual workday starts can shift the reset to land earlier in your working hours.
- Turning a process you've already solved into a reusable, named skill means an agent doesn't have to reason its way through the same steps from scratch every time.
- Standing personal preferences belong in account-level settings; project-specific instructions belong in a file the agent reads automatically at the start of every task.
06Rules 6-7: steer mid-run, turn off unused tools
- Saved instructions still cost context on every single run, so periodically auditing them for duplication or staleness keeps them from quietly becoming their own tax.
- Correcting a run while it's still working, instead of waiting for it to finish, stops you from paying for an entire output you already know is wrong.
- Every tool, plugin, or integration left switched on costs usage whether or not it gets used in a given session, so unused ones should be turned off by default.
07Rules 8-9: shorter reports, shorter answers
- Long tool-call reports that describe everything that went right, not just what went wrong, get re-read by the model and quietly inflate cost with no benefit.
- Asking for a short, direct answer by default and requesting full detail only when something breaks keeps output tokens proportional to what the task actually needed.
08Rules 10-11: exact file paths, progress records
- Pointing an agent at the exact file where a problem lives, instead of asking it to search the whole codebase, removes an entire investigation's worth of token spend.
- A short running progress file, what's done, what failed, what's next, lets work resume from where it left off instead of re-discovering the same state after an interruption.
09Live demo: vanilla versus enhanced Codex
- A controlled side-by-side test, same bug, same starting conditions, is the only way to see whether a stack of small optimizations actually adds up to a meaningful saving.
Terms worth knowing.
- Astra
- The GPT-6 model version powering OpenAI's Codex coding agent as of this video, with different context, caching, and usage-limit behavior than earlier Codex models.
- Codex
- OpenAI's AI coding agent product that runs multi-step coding tasks inside a chat or terminal interface, billed against rolling usage limits rather than per token.
- Luna Max
- A cheaper OpenAI sub-agent model that Astra can delegate narrow, well-defined tasks to at a fraction of Astra's own token cost.
- AGENTS.md
- A project-level instructions file that Codex reads automatically at the start of every task, used to store standing preferences so they don't need to be repeated.
- Tool call
- A single action an AI agent takes, like opening a file or running a command, that returns a report the model must read before deciding its next step.
- Usage budget
- A percentage of a usage allowance, such as a weekly limit, given to a model as an explicit target to stay within while completing a task.
Things they pointed at.
Lines you could clip.
00:00
“I reduced my codex usage by up to 91.75% and I did this by making GPT-6 Astra run hundreds of tiny experiments to find ways to optimize the way I was using codex.”
01:47
“Hey, complete this task in 3% of my weekly usage limit and Astra will try its best to stay within that budget.”
05:15
“With these optimizations, I was able to get 300 digital products made from the same allowance that initially would have gotten me only 25.”
16:40
“As you can see, we got a total of a 33.3% saving.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
I reduced my codex usage by up to 91 .75 % and I did this by making GPT -6 Astra run hundreds of tiny experiments to find ways to optimize the way I was using codex and what I found might actually surprise you. I used to hit codex usage limits every single day. However, after implementing a few practical rules for how I use codex, I haven't hit my limits in over a week.
so without further ado here are 11 things i changed that stopped me from hitting codex's usage limits and at the end i'll even do a live demo so you can see just how crazy these savings actually are oh and if we haven't met yet i'm doobie i build apps every day using codex and i've generated over fifty thousand dollars in 80 days using these ai tools so the very first thing i need you to do in order to make your codex usage last longer is to throw everything you know about ai out the window no i'm actually serious because With the rollout of Astra, OpenAI has made tons of tiny tweaks that make a lot of your previous token saving tips completely useless.
So let's play a game. Here I have a bunch of popular token optimization methods and I want you to guess how many of them still apply for GPT -6 Astra. Now, if you guessed three, you'd be completely wrong.
The answer is none of them. None of these apply to Astra. Embarrassing.
keeping context under 270 000 tokens is not even necessary anymore catch resets every five minutes well now they've extended it to 30 minutes oh and don't worry about changing effort levels in the middle of your conversations because codex won't reread your entire chat history anymore and the point i'm trying to make here is that as ai evolves we also need to adapt and create new methods that work for the next generation of models so let me start with rule one give Astra a usage budget.
With the rollout of Astra, OpenAI released a very important update that now lets your models see their own usage limits, meaning Astra can actually check how much usage you've got at any given moment. So now, unlike before, you can actually budget. Hey, complete this task in 3 % of my weekly usage limit and Astra will try its best to stay within that budget.
And I've found even just mentioning token budgets will make Astra work much more conservatively, but here's where it actually - gets really interesting. Because usage limits are now visible to the model, prompts such as run experiments to find out how we can reduce this workflow down to 1 % are actually viable now.
And OpenAI's own documentation supports this, saying Astra exhibits behavior patterns that can be optimized through prompting the model for your use case. Now, you may need to give up a little bit of your usage up front, but for repeatable workflows like reading emails and drafting scripts, you could easily reduce your token usage by 80 to 90%.
Just to give you an example of how effective this is, I'm filming another video right now where I use Codex to run a business selling digital products. The first product we created used 4 % of my weekly allowance. So I gave Astra some room to experiment.
I taught her to optimize the workflow with a goal of getting a digital product made for less than 1%. And Astra was able to absolutely smash that goal and get it down to three products made for only 1 % total of my usage limits. With these optimizations, I was able to get 300 digital products made from the same allowance that initially would have gotten me only 25.
That is the power of asking Astra to optimize the way it works. If you're finding the video helpful so far, be sure to give it a like. I make videos like this every week, so subscribe if you wanna see more.
Rule two, let Astra manage cheaper agents. Now, rule two is extremely important because unlike previous models, GPT -6 Astra specifically has been trained to be able to divide and delegate work to sub -agents that work in parallel, meaning OpenAI expects you to utilize sub -agents when working with Astra. Because the truth is you don't need Astra to personally execute every part of your plan.
I had Astra orchestrating Luna Max agents and Luna's token rates are around 98 % lower than Astra's. OpenAI actually cut Luna's prices by 80 % recently, which makes it seriously worth trying for routine work. Benchmarks even show that Luna Max is equivalent to Sonnet 5 at max, but for one seventh the price.
The trick with Luna models is to give it a very narrow job. Say you want AI to help manage your inbox. Instead of one model handling the whole process like Astra, you break it up into steps.
For example, one Luna agent could sort the emails and flag what needs attention. Another could draft the replies. Then you have Astra review the responses and handle anything complicated.
And I'm not telling you to manage all three agents manually. Nah, just tell Astra to manage Luna Max sub -agents for narrow execution and review at the end. That's literally it.
If Astra ends up redoing everything, then the task was clearly too big and you'll need to actually break it down further. Rule three, quick cheat code. If you're on the $20 plan, set a cron job to literally just ping Codex with one tiny message at 6am.
While you're still in bed, your 5 hour window starts then, not when you sit down to work. So let's say you work at 9am and you hit your limit at 10am. Well, your reset lands at 11.
Rule 4. Turn repeated mechanics into reusable skills. Once you've found a process that works, save it.
Don't make Codex figure out how to do the same job from scratch every single time. With your digital products, we kept repeating the same steps. Package the artwork, create the previews, build the download and check the files.
So we turned these steps into a reusable workflow. You can do this with a weekly report, a client proposal, your content research, whatever you regularly create. Just tell Codex, turn this successful workflow into a reusable skill.
The next time you need that job done, just pointed at the skill. And unlike the old days, skills are actually cheap to keep around now.
At the start of a task, Codex only reads each skill's name and a one -line description. It opens the full instructions when it decides to actually use that skill. So a skill you don't need today costs you almost nothing you only really ever need to iterate and go through a workflow manually once with your ai agent then every time after that the skill can explain it rule five save your preferences once and stop introducing yourself to codex every single time you open a new task for example act as a professional whatever keep it short where on next .js that's all token burned on repeat and usage spent teaching it the same thing again there are two places to save this to save instructions universally go to settings personalization and custom instructions here you can save whatever you want i like to add notes around my writing style and how i want answers formatted.
For my terminal users this updates your personal agents .md file and Codex reads that at the start of every task. If you're looking to add instructions that are more project specific just ask Codex to do it. Save my writing preferences in this project's agents .md so future tasks use them.
It will then create and edit that file by itself and if you keep correcting the same thing save that correction. Otherwise, you're paying to teach it again in every new chat. Now, it's also worth noting saved instructions still take up context on every task.
So keep them short. The savings come from skipping the repetitive prompting. But if your instructions are bulky and unused, you just lose unnecessary tokens again.
So to avoid things like that happening every month or so, just ask Codex, review myagents .md and skills, find anything duplicated, outdated or making simple tasks unnecessarily complicated. Show me what you'd change. Rule six, correct it while it's still working.
If Astra is building the wrong thing, don't just sit there and let it finish. Go to settings, general, follow up behavior. You got two options here.
A message you send mid run can steer the current run or you wait for the next one, which is queuing the message, the default. Pick steer. When you type no mid prompt, Codex gets the correction while it's still working.
Instead of finishing an entire job, you're about to throw away. It doesn't refund the tokens already spent, but it does stop you from wasting more tokens on a bad run. Rule 7.
Turn off the features you're not using. OpenAI's pricing clearly states model choice, context, reasoning, tool use, retrieval, and catching all affect usage. So every tool Codex has switched on costs you something whether you use it or not.
Start with plugins. I find myself installing plugins for one -off use cases and I always forget to turn them off. It's good to go here and just turn things off.
You might be surprised at how many random things you've got here. And also watch image generation as well.
It's the same usage pool as your chat, but burns through your limits three to five times faster than a normal prompt so don't ask codex for random generations now if you're not sure what switched on type slash status in a fresh chat it shows your context usage and your rate limits if a chunk of your context is gone before you've typed a word that's instructions tools and plugins it won't tell you directly which plugins are eating up your context so switch things off one at a time and see what moves the needle the rule of thumb is if you didn't turn on intentionally for this job turn it off okay if you've somehow made it this far into the video then you're in luck because i'm looking to grow a small exclusive community of people where i teach you everything i'm learning about ai from the basics up to growing and launching your own apps now i'm not really announcing this anyway i'm just going to keep hiding it at the end of my videos because i only want people who one actually have an attention span and two are genuinely motivated to build and launch something so if that sounds like you i'll leave a link below
Rule 8. Stop making it read every successful check. Every time Astra does part of a task, it makes what's called a tool call.
Opening a file is a tool call. Running your website to see if it works is a tool call. Searching the web is a tool call.
And every tool call sends a report. back to Astra it then has to read the whole report before it can take the next step but here's the problem those reports are long by default say you ask Astra to look through 200 files for a spelling mistake it doesn't get back a note saying found two it gets back a full write -up on every single file including the 198 that were fine and it reads all of it and you're paying for it imagine you hired someone to do that job and they came back and read you a full report on every file that had nothing wrong with it so tell the same thing.
Keep the reports from the tool calls short. When something works, just say it worked. Only give me the details when something goes wrong.
We tested this across six matched comparisons. The same job with shorter reports used about 6 % fewer tokens. And this can compound pretty quickly.
Everything still got done and the problem still showed up. Astro just stopped reading pages that marked if something was fine or not. And again, if you don't want to say it every time, tell Codex to make it permanent.
Now it's worth noting, be careful because if you cut it too short, it can miss the one line that... explains what broke. So try it on a real job before you use it.
for production rule nine stop paying for essays you didn't ask for every word astra generates adds output tokens in english a token is roughly three quarters of a word so a hundred word response is around 130 tokens now that's fine when you actually need the explanation but if you asked it to change a button you probably don't need six paragraphs about its incredible button changing journey now there are plenty of ways to reduce astra's skills like caveman which teach it to talk like caveman which is also kind of entertaining or my personal favorite is I have ADHD which forces Astro to speak more concisely with line breaks.
If you don't want to install any skills OpenAI also recommends adding this prompt to your agents .md file which should also do the trick. The rule here is to make the response as short as the job allows because this will cut out the output portion of your usage. It also saves you from Codex finishes something and when you want the full explanation you can just ask for it rule 10 stop making astra guess where to start if something's broken on your website don't just say debug my website tell it which page is broken what happens when you click the button and what should actually happen instead if you know which file controls that page give it the file people don't realize how much of their usage gets wasted from the model just searching for stuff we actually tested this ourselves across six comparisons giving astra the correct file path used around 23 fewer total tokens a few extra words in your prompt can save an entire investigation.
And if you don't know where in your code base it is, it can be useful to have an agent map out your code base and create a skill so future bug hunts can find the problem faster. Rule 11, make it leave a record of what's finished. Nothing is more frustrating than watching AI spend your usage doing something it already did.
So for longer jobs, have Astra keep a short progress file, what's finished, where the files are, what failed and what needs to happen next. For example, if it's drafting emails, record which... conversations already have drafts then if something crashes or you come back tomorrow it has a clear starting point in my current workflow saved assets and draft ideas let me resume after interruptions you've already paid for that work so make sure the next attempt can pick up from it all right guys as promised just to prove to you that i'm not lapping here i have vibe coded a dashboard for vanilla codecs as well as the enhanced codecs which is going to include all of our rules and optimizations that we talked about in this video and we're going to be able to see how much tokens they used to fix this bug so what is the bug so basically we have this website here and you can click through these buttons to change the background however we also have a dummy website here and these buttons don't work so the bug is really simple fix the buttons so they work
So we're starting off with Vanilla Codex over here. We've gone ahead and sent it the prompt. Let's see how it performs.
We're not telling it where to find the problem. We're not telling it to use our sub -agents. We're not turning off our plugins.
This is completely vanilla how someone would normally prompt their Codex agent. As you can see over here in our dashboard, the tokens are going up for Vanilla Codex. Okay, boom, looks like it's fixed.
Let's go ahead and have a look.
And yep, it seems to be working. Nice. Looks good.
Now for the enhanced codec session, we're going to be sending it a much more specific prompt, telling it exactly where the components are, giving it a budget to follow, telling it to delegate to LunaMax sub -agents. We're also going to be turning off a lot of our plugins that are unnecessary. So all of these are going off.
Also, rather than just giving it the caveman skill, I've just told it to talk like a caveman. So that should also do the trick. So again, we have a faulty website and the enhanced codex agent is off.
Let's see how it does. Now it is worth noting when you are using Lunamax agents, it does go a little bit slower. So I am expecting it to be slower than Vanilla Codex, to be honest.
This is about saving your tokens, not spending them as fast as possible. As you can see here, we have one Lunamax sub -agent running right now. All right, Enhanced Codex is now complete.
So let's go ahead and refresh the page. Let's see. And yeah, boom, working just as expected.
Now let's have a look at the results. Boom. As you can see, we got a total of a 33 .3 % saving.
Now it's worth noting because the debugging task itself was quite small it's not going to be like a massive difference like you're not going to see a 90 difference here but i guarantee that if you actually give this a long -running task if you actually give this a much harder debugging tool or you're building an app this saving is going to be much much higher and there you have it 11 practical ways to never hit your codex limits again like the video if it helped and check this video out youtube thinks you'll love it
The Hook
The bait, then the rug-pull.
A claimed 91.75% cut in Codex usage opens the video, and instead of just asserting it, the creator spends 16 minutes walking through 11 specific changes and then proving the stack of them with a live side-by-side test.
Named ideas worth stealing.
01:45list
11 Rules for Cutting Codex Usage
- Give Astra a usage budget
- Let Astra manage cheaper sub-agents
- Ping a cron job before your workday starts
- Turn repeated mechanics into reusable skills
- Save preferences once, at the right scope
- Correct it while it's still working
- Turn off features you're not using
- Stop making it read every successful check
- Stop paying for essays you didn't ask for
- Give it a starting point instead of a guess
- Make it leave a record of what's finished
The video's full listicle structure, presented as a direct replacement for token-saving habits that no longer work on GPT-6 Astra.
Steal forany AI-agent-heavy workflow that's billed against a usage limit rather than a per-token price
How they asked for the click.
VERBAL ASK
09:21product
“if that sounds like you I'll leave a link below”
A soft, low-pressure pitch for a paid Skool community dropped mid-video, framed as 'not really announcing this' rather than a hard sell.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
