16:54Mark Kashef · Tutorial
GPT-6 Astra's Computer Use Is Ridiculously Good
Five ways Mark Kashef points Codex's computer use at apps that have no API, from flight searches to phone settings.
September 5thThe argument in one line.
Across seven parallel runs of the same research-and-build assignment, GPT-6 Astra's reasoning-effort setting changed completion time and token spend unpredictably but didn't reliably improve output quality, so Medium is the sane default and higher tiers should be reserved for a specific gap Medium leaves open.
Read if. Skip if.
READ IF YOU ARE…
- You use Codex or a similar coding agent regularly and want a data point on whether raising the reasoning-effort setting is worth the extra wait.
- You're deciding between GPT-6 Astra and the older Sol model for day-to-day building tasks.
- You want a repeatable method for running side-by-side AI settings comparisons on your own prompts.
SKIP IF…
- You're looking for a technical explanation of how reasoning-effort settings work internally; this is an applied usage test, not a mechanism breakdown.
- You need results on a pure coding task; this test used a research-and-build business assignment, not raw code generation.
The full version, fast.
The video tests whether GPT-6 Astra's reasoning-effort setting, Low through Ultra, actually changes output quality, using seven identical research-and-build assignments run in parallel Codex threads, plus GPT-5.6 Sol on High as a control. Each run had to research a SaaS opportunity on Reddit and X, define a moat, sketch an Excalidraw plan, and ship a working prototype. Completion time ranged from 30 to 46 minutes and token spend from 6.7 million to 21.73 million, with no consistent relationship between effort level and either speed or output quality; Ultra's high token count came mostly from three subagents. The creator's conclusion: start every assignment on Medium, since it finished fastest with a working core flow, and only raise effort when a specific gap in the result demands it.
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 – 00:57
01 · Does Astra's effort setting matter?
Cold open framing the central question: does raising GPT-6 Astra's effort level actually improve results, and does it stop the model from being 'lazy'.
00:57 – 02:02
02 · Astra Low vs Sol High and the laziness debate
Breaks down Tibo's viral X claim that Astra on Low beats Sol on High, and OpenAI's own documentation on why Astra asks more clarifying questions.
02:02 – 02:53
03 · The experiment setup
Lays out the six controlled variables: same context, same access, same build size, same conditions, Ultra kept separate, and full token/time tracking.
02:53 – 05:00
04 · The research and build assignment
Walks through the shared prompt section by section: customer, evidence, business/moat, plan, product, and delivery requirements.
05:00 – 06:06
05 · Launching Codex tasks at different effort levels
Shows the prompting trick for spinning up multiple renamed Codex threads at different effort levels to run in parallel.
06:06 – 07:21
06 · Astra Low results
Astra Low built Clearshift, a cleaning correction and recheck tool, in 37:37 using 14.81M tokens across 8 Reddit threads and 161 canvas elements.
07:21 – 08:31
07 · Astra Medium results
Astra Medium built Scopewell, a cleaning scope and pricing tool, in 30:41 using 10.23M tokens, the fastest run of the whole test.
08:31 – 09:45
08 · Astra High results
Astra High built Fieldwork, a cleaning-schedule recovery tool, in 41:21 using 14.33M tokens and produced the most detailed canvas so far, at 186 elements.
09:45 – 11:40
09 · Astra Extra High results
Astra Extra High built Scopekeep, a remodeling extras and billing tool, in 45:14 using 12.29M tokens; visually similar to High but judged slightly worse underneath.
11:40 – 12:55
10 · Astra Max results
Astra Max built Fieldnote, a remodeling scope and approval tool, in 46:10, the longest run, using only 10.55M tokens and still no X evidence.
12:55 – 13:59
11 · Astra Ultra and subagents
Astra Ultra built Accord, a remodeling-additions agreement tool, in 42:10 using 21.73M tokens by spinning up three subagents that alone added 7.12M tokens.
13:59 – 15:24
12 · GPT-5.6 Sol High comparison
The Sol High control built RelayOps, a field-service recovery tool, in 32:49 using just 6.70M tokens, far less than any Astra tier, though its website looked the roughest.
15:24 – 16:23
13 · Which effort level I'd use
States the verdict: start on Medium since it finished fastest with a working core flow, and only escalate effort when a concrete gap in the output requires it.
16:23 – 17:09
14 · Final takeaway and the free Astra Super Guide
Wraps with the conclusion that Max, Extra High, and Ultra rarely earn their extra time or tokens, and points to a free downloadable Astra Effort Guide.
Lines worth screenshotting.
- GPT-6 Astra on Medium finished a full research-and-build assignment in 30 minutes 41 seconds, the fastest of seven effort levels tested, using 10.23 million tokens.
- Astra Ultra took 42 minutes and burned 21.73 million tokens, nearly double most other effort levels, largely because it spun up three subagents that alone added 7.12 million tokens.
- Token spend across Astra's effort levels was not linear: High used 14.33 million tokens while Extra High, one tier up, used only 12.29 million.
- GPT-5.6 Sol on High completed the same assignment in 32 minutes 49 seconds using just 6.70 million tokens, less than a third of what several Astra tiers consumed.
- None of the seven runs pulled meaningful evidence from X; only Astra High and Sol High cited any X threads at all, and even those cited just one or two.
- Astra Low and Astra Medium each independently converged on a commercial-cleaning SaaS idea despite the prompt never naming an industry.
- Higher effort levels didn't reliably mean more source research: Astra Low read 8 Reddit threads, more than Medium's 6 and on par with several higher tiers.
- You can spin up multiple Codex threads from one prompt by asking the model to create named threads, set a different effort level on each, and run them all at once.
- The creator's standing daily preference is Astra on Medium with Fast mode, independent of this specific test's results.
- Every result card in the test exposed an unresolved gap, from missing X evidence to an unverified native download, meaning no single effort level shipped a fully finished deliverable.
- Astra Extra High's interface looked similar to High's on the surface but was judged worse underneath, despite spending more wall-clock time to get there.
- GPT-6 Astra's own documentation states it is designed to ask more clarifying questions and may stop when it expects the user to make a reasonable assumption instead of proceeding.
What actually changes when you raise Astra's effort level
WHAT TO LEARN
Across seven identical assignments, effort level moved completion time and token spend unpredictably in both directions but rarely made the output meaningfully better, so Medium is the sane default and higher tiers are for solving a specific problem, not a default upgrade.
02Astra Low vs Sol High and the laziness debate
- OpenAI's own documentation says Astra is designed to ask more clarifying questions and can stop when it expects you to make a reasonable assumption instead of just proceeding.
- A viral claim that Astra on Low beats Sol on High is what triggered this test, but the claim conflates raw capability with how much effort a specific setting spends.
03The experiment setup
- Fair AI comparisons need explicit controls: identical starting context, identical tool access, a fixed build scope, and any condition that changes the workflow (like Ultra's subagents) tracked separately.
- Recording both time and token spend for every run is what exposed that more effort doesn't reliably cost more tokens.
04The research and build assignment
- Naming a customer segment without naming the industry is a way to test whether a model defaults to a narrow vertical on its own.
- Explicitly telling the model that competitors can reproduce its software in a weekend forces it to argue for a real moat instead of assuming code alone is a defensible business.
- Splitting an assignment into customer, evidence, business, plan, product, and delivery sections gives every configuration room to show a difference, if one exists.
05Launching Codex tasks at different effort levels
- You can ask Codex itself to create multiple named threads, each pinned to a different effort level, and run them on the same prompt at the same time.
- Running comparisons this way removes hand-configuring each thread individually as a bottleneck to testing settings at scale.
06Astra Low results
- Astra Low finished in 37:37 using 14.81M tokens across 8 Reddit threads, producing 161 canvas elements, but pulled no evidence from X at all.
- Low's proposed idea worked but broke no new ground, and it still required a recheck before the flow could be called finished.
07Astra Medium results
- Astra Medium finished in 30:41 using 10.23M tokens, the fastest and leanest run in the entire test, with a price and approval flow that survived a reload.
- Medium converged on the same industry as Low without being told to, suggesting the model's default reasoning path favors this vertical for this kind of prompt.
08Astra High results
- Astra High produced the most detailed canvas up to that point, at 186 elements, and was the first run to cite an X thread alongside its Reddit sources.
- High's business idea was judged the clearest one-line pitch of any run so far, showing detail and clarity don't always track token spend.
09Astra Extra High results
- Extra High took longer than High but used fewer tokens, showing effort level doesn't scale token spend predictably.
- Extra High's interface looked similar to High's on the surface but was judged worse in navigation and layout once examined closely, despite the extra time spent.
10Astra Max results
- Astra Max took the longest wall-clock time of any run but used the second-fewest tokens, decoupling time spent from tokens spent.
- Even at this high a tier, Max still skipped X evidence entirely, repeating the same gap seen in several lower tiers.
11Astra Ultra and subagents
- Ultra used by far the most tokens of any run because it delegated to three subagents, which alone accounted for roughly a third of that total.
- More tokens and more delegation did not translate into a visibly better final product than Medium or High produced with a fraction of the spend.
12GPT-5.6 Sol High comparison
- The older Sol model on High finished in 32:49 using only 6.70M tokens, under a third of several Astra tiers, making it the most token-efficient run in the test.
- Sol's output covered more source threads than most Astra runs, but its actual website was rated the ugliest of the seven.
13Which effort level I'd use
- The recommendation is to default to Astra Medium for new work since it finished fastest with a working core flow at the lowest token cost among the Astra tiers tested.
- Raise effort above Medium only when a specific, identified gap in the output calls for it, not as a default assumption that more effort means a better result.
Terms worth knowing.
- Effort level
- GPT-6 Astra's reasoning-effort setting, adjustable from Low to Ultra, that controls how much the model reasons before acting.
- Sol
- GPT-5.6 Sol, OpenAI's prior model, used here on High as a control group against Astra's six effort tiers.
- Subagent
- A delegated sub-task thread a model spins up to parallelize work; only Astra Ultra used them in this test, and they added significant token cost.
- Moat
- A claimed sustainable advantage a proposed business has over competitors who could otherwise copy it quickly.
- Excalidraw
- A whiteboard and diagramming tool the assignment required each model to use to sketch its own visual business plan.
- Canvas elements
- The count of individual shapes, notes, and diagrams on an AI-generated Excalidraw board, used here as a rough proxy for output volume, not quality.
Things they pointed at.
Lines you could clip.
14:15
“You should not bring a nuclear bomb to a fistfight all the time.”
03:32
“Assume customers and competitors can use Astra to reproduce competent software quickly, meaning don't build a SaaS that can be vibe coded in a weekend.”
16:05
“I see very few cases where going on max extra high or ultra is ever worth the extra time effort or tokens.”
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
So GPT -6 Astra has been out for a bit now, and one of the biggest points of contention on both YouTube and X is the following question. What's the right effort level to get the most value or bang for your buck from this model? I've also seen people say that this model is inherently lazy.
It does what you tell it to do, but nothing more, and sometimes not even that. So I ran a few experiments to see if changing that effort level also changed that behavior. I tested every Astra level on the exact same research and build project.
And I even threw in soul as a comparison. So in this video, I'm going to walk you through what each one came back with, how long it took to process, how many tokens they burned and whether effort had any influence whatsoever. By the end, you'll be able to make a solid judgment as to which effort level might fit your day -to -day projects.
Let's dive in. So this post on X is the very reason why I wanted to run this test. Thibaut is OpenAI's core products lead, and he's also known as the guy that keeps resetting your usage seemingly every hour.
And he made this bold claim. GPT -6 Astra on low performs better than GPT -6 Sol on high. So Sol is the prior model.
And it's a model that is now being forgotten about because of this new shiny toy. But for many purposes and tasks, it's still more than sufficient to execute. Now when you do some more digging into actual user experience and anecdotes from using this model, especially on low, you hear things like it's stopping way too early and asking for permission every five minutes.
So I wanted to gauge two main things from this experiment. One, does extra effort necessarily give you better results? And two, does it give it more autonomy to stop asking you for permission and stop being apparently lazy?
If you go to OpenAI's official documentation, it says the following. The model is designed to be a more effective collaborator and thus more likely ask questions of the user. And then it says this.
This can cause it to stop when the user may expect it to make reasonable assumptions. So where something like Fable 5 .1 is really good at, where you give it the general understanding, the TLDR of where you're going, and it's able to not only extrapolate, but also thoughtfully add additional steps that it sees fit, it seems like Astra, at least at some effort levels, might not do this as effectively.
So to make sure that this experiment I'm about to walk you through was as transparent and equal as possible. I controlled for these six variables. We kept the same context, the same access to tools, MCPs and skills, same scope of the build as well as same conditions.
And we separated ultra because many times ultra spins up a series of sub agents. You'll see exactly the little trick that I use to actually spin up a series of threads with different effort levels using the same prompt. It'll be really cool.
And last but not least, we kept track of how many tokens and how much time each one of these threads took. Now, this is the prompt that I sent to all seven threads, and I'm going to break it down section by section. The crux of the task was to give every agent the ability to go and research everything it could find on X and on Reddit, which are two great places to listen for user feedback.
This was the core customer that it was optimizing for. So find a worthwhile SaaS opportunity for owners of a small service business between two and 20 employees. I haven't chosen the business category, the problem, or the product.
Start with their problems, consider three opportunities, then choose one. Then for the evidence portion, we just asked it to go on these platforms, like I mentioned, and look into three existing alternatives to its initial idea. And the goal was to give a short list of all of the repeated customer frustrations, workarounds, spending signals, and counter evidence.
Now, when it comes to assessing the business opportunity, this is the most important line in the whole prompt. Assume customers and competitors can use Astra to reproduce competent software quickly, meaning don't build a SaaS that can be vibe coded in a weekend. There should ideally be some form of moat to that specific business model.
For the planning portion, I wanted it to be able to create its own Excalibur files by using AI to create JSON files that are representative of doing this. Because now with these more advanced models, I no longer need a skill. I can tell it, go make me an Excalibur file.
And Excalibur is basically this file that allows you to annotate, create diagrams as if I've doodled them myself. Instead of using a skill, I wanted each level to demo creating the entire file itself to see how well organized was the thought pattern for that specific effort level. For the main product, I wanted it to build a polished website, very similar to how I put together these websites for you on YouTube to explain these concepts.
I wanted it to create a ChatGPT site that breaks down its entire plan to build said product after the research. And last but not least, on top of the website and on top of the Excalibur file, I also wanted to create a working prototype of what that SaaS would look like. So by breaking up the task in these different sections, it would give enough breathing room for each effort level to shine if there is a real difference between them.
And the core things I wanted them to answer was, one, can we trace the claim as to why this is a good business idea? Why would anyone pay for this? Do the pieces actually connect when it comes to the avatar you pick, the pain point that you say exists and where you think it exists, the product?
Does the workflow work? Does the SaaS make sense to exist as a SaaS product? And finally, From the follow -through perspective, did it finish and verify its own work?
Did it actually ruminate and think through what happened? Now, these are all the threads that we ran. And I'd be lying to you if I told you that I spun up all of these individually by hand.
All you have to do is actually ask Codex to spin up a series of threads and tell it to, one, rename them based on the effort level, and two, switch the effort level, and three, pin it within a certain folder. All you have to do is prompt something like this. I want you to take the exact same prompt.
You can come up with it. I want it to be a very simple task. And I want you to create three physical threads.
Rename them something that's easy to read on the sidebar. And I want you to have each one at a different level for Astra. One on low, one on medium, and one on high.
And I want you to run this prompt at the exact same time right now and make sure that you monitor the progress of each one. So if you send this over, This is a newer thing that's working both in Claude and Codex.
It should be able to spin up. If we click on Recents right here, brand new threads. You'll see the test threads will start to split out.
There we go. And now we have Astra, Medium, and Low, and High. And they're all running, ready to go.
If you click into them, you'll see the exact same task in there. And this is how you can start to do testing at scale. And the result of every one of these threads was a canvas file.
Like we said, it looks like this, some form of website and a full breakdown as well as our statistics as to how many tokens it spent the time. And what was the core result? So Astra on low took 37 minutes spent almost 15 million tokens went through eight different threads on Reddit, not X.
Then it had 161 canvas elements. Now this is not necessarily a proxy for how good it is, but it's just how much work it did. It came up with this product tracks, missed cleaning tasks through correction and recheck for commercial cleaning owners, managing recurring office sites, proposed moat client specific standards in a history of fixes that worked.
Very interesting. The website looks something like this. Again, we're not assessing the beauty of it.
We're assessing the core concept. But this is its draft for this apparent SaaS. Not too impressive, but could be a business for all I know.
The one big gap here is that there is no evidence drawn from X, and it didn't do anything above and beyond. Now, Astra on Medium only took 30 minutes, 10 million tokens, went through six Reddit threads, and this was the product they put together. Prices extra cleaning work and records client acceptance for commercial cleaning owners with written service agreements.
And the proposed moat is accurate scope records and estimates improved by actual jobs. Not sure how much of a moat that is. The one interesting thing you'll find is that this is once again in the cleaning industry, even though I didn't lay that out in the prompt itself.
So given the avatar and the scope I gave it, maybe given everything that was mentioned in terms of a condition, this seems to be the best industry. The website looks similar. If you click on explore plan and we go on something like Canvas, it just seems to have a lot more of a clean breakdown as to what the customer looks like, the alternatives, the opportunity, the owner journey.
It has a decent breakdown of all the consequences. And I'd have to read through this to really get down to the nuts and bolts as to how good of a business model this is. But in terms of the quality and the variety of different analysis, it's definitely better than low.
And it took less time and tokens to do so. Now, as we go to Astra High, it spends close to 40 minutes and then spends 14 million tokens, goes through Reddit, and finally goes through one post or one set of posts on X. And then it has the most detailed canvas, and this is the business it came up with.
Reassigns visits when a cleaning crew cannot work. That's probably the most simple one -liner I've seen so far. Four residential cleaning owners juggling crews and appointments.
Proposed moat. Company -specific constraints and outcomes that improve scheduling. Now of all the websites so far, this is the cleanest and the clearest in terms of what is the overall vision and what is the pain point.
And the pain points here, 7 42 AM Cedars van won't start. So there's two recurring visits. Someone needs to cover.
We need to be able to fill a slot for what was going to be a cancellation. So this idea is trying to minimize unrealized revenue. Then the big picture question here is, is it worth it to have spent 40 minutes and 14 million tokens?
Or could you have spent the 30 minutes on medium and just given it a more dialed in prompt or a followup prompt to get to this exact same point. Next up we have Astra extra high 45 minutes, but it actually spent a little bit less than high, which is 12 million. And what's really interesting is how I know this is not AGI and how a lot of these claims to me are very superfluous is.
Every single effort level should be a proxy for efficiency in a way, but it's not. You see high, close to 14 million tokens. You saw low, also close to 15 million tokens.
Then you saw medium, close to 10 million tokens. There is variation across the board and it's very unpredictable. And if I were to run this again, I would tell you that most likely the time to take each one will be the same.
But in terms of tokens spent, it is increasingly unpredictable. Now the business model here isn't related to cleaning for once. It's taking remodeling extras from approval into billing.
For small remodeling owners losing track of chargeable extras, and the moat is tailored scope checks informed by what it got approved, billed, and paid. Bit of a mouthful. Now in terms of the website, it looks on par with Astra on high, but it's actually worse when you go underneath the hood.
So it has a similar layout. It has these similar four tabs. And then as you go to the connected canvas, it does open it in a brand new tab, which is a nice little effect.
You can filter and click on little buttons here to help you navigate, which is cool. But in terms of the scope, the look and feel. of this canvas, it looks slightly worse than the high version.
This was the high version. It also has buttons, just a different placement, slightly more squeezed. And there's even more proof for every single one of these claims.
So when you go to customer evidence and you click on this, it goes directly to that associated thread. So not only do we have a better breakdown and even more sources for every single element on the screen that links to the exact threads that it sourced it from. But if you compare them head to head, this is a plain Jane looking layout.
It's less detailed. It has less references and the navigation is slightly better, but outside of that, there's no added benefit for the extra five, 10 minutes and possible tokens that you're spending on top of medium and high. Now Astromax was confusing for me because it did spend 46 minutes, but spent only 10 million tokens in that time.
So it was efficient, but slightly inefficient in terms of the time it took to get to those number of tokens. The business idea was to keep an extra job scope price and approval in one record. So I believe it's database oriented or data hygiene oriented for residential remodeler owners with teams.
of two to 10 people. And then a proposed moat, an adopted job suite routine, assisted setup and bookkeeper referrals. This is interesting.
So I was thinking about a meta way to have business based on referrals. One thing I want to point out here is despite being on max, you don't have any reference to X, even though we saw that pop up in the high version. So once again, more evidence that not necessarily getting better results, spending more tokens, or in this case, spending more times to spend said tokens.
Now in terms of the website, it does look cleaner. It looks more polished. It actually looks like a real product compared to high and extra high, but This overall, it does sell me the vision of what it is and what problem we're trying to solve.
But would you spend an extra 10 minutes for this versus adding an additional prompt or a couple extra prompts with a lower tier model? And then we get to the big daddy Astra Ultra, where this one took 42 minutes, so less than max, but it spent a whopping 22 million tokens to get to this result. This is what you get with 22 million tokens.
Records and agrees remodeling additions before work starts. So basically, And in terms of the website, it doesn't look like a DaVinci masterpiece.
It looks very similar to everything else we saw with slight spacing and extra polish, but it's not polished that is additive to the experience. So if you click through the opportunity, again, you have a nice little embossed button. You have a canvas that looks almost identical to the exact same thing we saw before.
We have a weird placement of two sources, so now you're getting an extra set of sources per bubble. Is that worth the extra time and tokens, 20 million tokens, and sub -agents that were executed to spin this up? Now, as our control group, we also tried Sol High, because if you go back to Tebow's comment, you're supposed to be able to use Astra on low and have an equivalent experience as Sol on high.
This one took only 32 minutes, only 6 million tokens, so if you've forgotten about Sol, there's a reason to remember it. Astra shouldn't be used for everything.
And then... Surprisingly, this one went through six Reddit, two X threads. So again, quantity is not necessarily a proxy for quality, but just the idea that it at least checked them as sources.
And then if you go to the business plan, plans recovery when delays disrupt a service route, very similar to what we saw before. Owner dispatchers of recurring service firms with two to 20 staff. And the moat is recovery outcomes, maintained rules, and accountable support.
So if you go to the website, I expect this to be not as pretty, but it should be functional. All right, so this is unbelievably ugly. And it's so bad that I can't really look at it without wanting to close my eyes or at least squint my eyes.
But I think it's attempting on the first image to show me a diagram of what the business model is. But I can say, you know, very easily. Astra is way better at websites and creating that quick ChatGPT site.
But the real question is, could you have just used the skill to make the pretty website with the proper spacing and still use Sol on high or medium? Now from using it in this experiment and using it actually in my day -to -day, I love using Astra on medium on fast mode. It's like my favorite model of every version of the Codex models.
It's fast, it doesn't consume tokens at an alarming rate, and it's pretty efficient. In this experiment, I found the exact same thing to be true, and hopefully you did too, where the website, the structure, the business plan, the business model, the canvas, all looked very similar between something like medium and high. If you wanted that extra bump in quality, then you could go high and spend an extra couple tokens.
But for most day -to -day tasks, if you want to use Astra, even though you shouldn't use it for everything, medium is more than sufficient for the majority of tasks. And that's pretty much it. So hopefully this gives you a good breakdown and a good understanding as to what effort level might be the perfect one for you.
Just based on this test and a series of others that I've done, I see very few cases where going on max extra high or ultra is ever worth the extra time effort or tokens. Now I've ran many more tests and I documented a lot of those findings along with additional things that I found in a few nuggets on when to use certain effort levels with Astra for certain types of tasks.
So if you want to check that out, I'm gonna make it available to you down in the second link in the description. You can just grab that for free, take it, throw it into your codex. into your cloud, and then just have it basically break it down for you.
And as always, if you want to be 10 steps ahead with anything AI, especially with applying it practically to your day to day, whether it's your job, your business, you're looking to learn or earn, check out the first thing down below. And I always go infinitely deeper on topics just like this inside my early adopters community.
For the rest of you, if you enjoyed this and found it helpful, I'd super appreciate a like on the video. It helps to reach and a comment if you so choose. I'll see you all in the next one.
The Hook
The bait, then the rug-pull.
Everyone arguing about GPT-6 Astra's effort slider is really arguing about one unanswered question: is it actually smarter on High, or just slower. Mark Kashef sends the identical research-and-build brief into seven parallel Codex runs, from Astra Low to Ultra plus a Sol High control, and lets the stat cards do the arguing.
Named ideas worth stealing.
02:09list
Six controlled test variables
- Same context
- Same access
- Same build size
- Same conditions
- Separate Ultra
- Keep the receipts
The six things the creator held constant across all seven effort-level runs so the comparison would be fair: identical starting files and briefs, identical tools/MCPs/skills, one user role and three main screens, the same recorded speed setting, Ultra pulled out as its own condition since it triggers subagent delegation, and every run's sources and artifacts logged.
Steal forany side-by-side AI model or settings comparison you want to run credibly
04:49list
What counts as finished (follow-through check)
- Can we trace the claim?
- Do the pieces connect?
- A clean finish, nothing left hanging
- Did it verify its own work?
The evaluation rubric applied to every run's output: whether the business claim traces back to real evidence, whether customer, journey, canvas, and implementation all agree with each other, and whether the model verified and wrapped up its own work instead of leaving it half-finished.
Steal forgrading any AI-generated business plan or prototype for internal consistency
How they asked for the click.
VERBAL ASK
16:36link
“Get the free Astra Super Guide through the second link above ... check out the first thing down below ... I always go infinitely deeper on topics just like this inside my early adopters community.”
Soft dual CTA delivered right after the verdict: a free lead-magnet guide (second description link) plus a paid community plug (first link), followed by a like/comment ask.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
