26:26Pat Simmons · Review
GPT-5.6 Soul vs Fable 5: I Gave Both Agents the Same Three Builds
Same prompts, three one-shot builds, two frontier coding agents — and one surprisingly clear winner.
July 11thThe argument in one line.
Given the exact same one-shot prompt and no revisions, GPT-6 Astra produced more detailed and more accurate builds than Claude Fable 5.1 across three different projects, largely because it could run unsupervised for far longer without hitting a usage limit.
Read if. Skip if.
READ IF YOU ARE…
- You're choosing between Claude Code and Codex/GPT-6 Astra for long, expensive one-shot builds and want real cost and time numbers instead of vibes.
- You're curious what an AI agent left completely alone for 11 hours actually does with that time.
- You want a sense of how far a single prompt can currently push 3D generation, historical research, or basic game mechanics.
SKIP IF…
- You're looking for a coding tutorial — this is a reaction video watching two models build unattended, not a walkthrough you can follow.
- You need a controlled, reproducible benchmark. This is one creator's subjective run of three builds, not a test suite.
The full version, fast.
A creator gives GPT-6 Astra (via Codex) and Claude Fable 5.1 (via Claude Code) the identical one-shot prompt, no revisions, on three builds: a retro iPod app with AI-generated 2000s-style music, a scrubbable 3D D-Day beach reconstruction, and a Nuketown-style FPS clone. Astra wins the iPod build faster and cheaper (about $13 versus $41), then spends nearly 11 unsupervised hours on the D-Day build cross-checking real 1944 tidal data and refining grass texture, versus Fable's 40-minute pass that rendered troops as plain boxes. Astra's FPS clone also has sharper character models. The through-line: Astra can run unattended far longer without hitting a usage limit on a $200-a-month plan, and the creator argues that patience is now as much the differentiator as raw model quality.
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 – 00:29
01 · Cold open: the rules of the head-to-head
Both models get the identical prompt, a single one-shot output, no revisions, deployed live to the internet. Three builds previewed: iPod app, D-Day reconstruction, Call of Duty clone.
00:29 – 02:23
02 · Build 1 setup: the retro iPod app
Both models get /goal, follow prompt.md, do not stop until complete. Spec: a native macOS iPod app loaded with ElevenLabs-generated songs styled like the early 2000s, credited as inspired by Pietro Schirano's build.
02:23 – 05:13
03 · Build 1 reveal: Fable's iPod
Fable 5.1 finishes in roughly 80 minutes. Clean UI with dents, scan lines, working scroll wheel with audible clicks, but auto-added unrequested extras (Projects, Now Reading pulling in real GitHub repos).
05:13 – 08:12
04 · Build 1 reveal: Astra's iPod + cost comparison
Astra finishes in about 37 minutes with sharper text and a more realistic dial, but no tactile sound effects. Usage panels show roughly $13 for Astra versus $41 for Fable.
08:12 – 12:13
05 · Build 2 setup: the D-Day reconstruction
Prompt (written by a research agent) asks for historically accurate terrain and a scrubbable minute-by-minute timeline, citing a similar Waterloo recreation as inspiration.
12:13 – 15:35
06 · Build 2 reveal: Fable's D-Day
Fable finishes in about 40 minutes with a dense, information-heavy UI (timeline, tide chart, plan-vs-actual view) but renders troops as plain boxes with thin 3D detail.
15:35 – 19:59
07 · Build 2 reveal: Astra's D-Day
Astra ran unsupervised for 10 hours 51 minutes, cross-checking real June 6, 1944 tidal data and refining grass texture. Result: a walkable village with modeled rubble and detailed troop models, without hitting any usage limit on the $200/month plan.
19:59 – 23:05
08 · Build 3 setup + Fable's Nuketown FPS
Prompt asks for a browser-based Call of Duty clone on Nuketown, telling the agent to research real FPS mechanics first. Both models take about 47 minutes. Fable's clone is playable with real gunplay but boxier art and some glitching.
23:05 – 26:22
09 · Astra's Nuketown FPS + final verdict
Astra's clone has more realistic character models and a map closer to the real Nuketown, though reload and scope are broken in both builds. Creator declares Astra the better coding model for the first time in this comparison series.
Lines worth screenshotting.
- A one-shot, no-revision native macOS iPod app took Claude Fable 5.1 about 80 minutes and roughly $41 in Claude Code usage, while GPT-6 Astra finished the same build in about 37 minutes for roughly $13.
- GPT-6 Astra spent 10 hours and 51 minutes one-shotting a 3D D-Day beach reconstruction, at one point pulling real tidal data from June 6, 1944 because its generated tide pattern looked unrealistic.
- Claude Fable 5.1 finished the same D-Day reconstruction in about 40 minutes but rendered troops as unlabeled boxes instead of soldiers.
- An 11-hour, unsupervised agent run didn't trip usage limits on a $200/month ChatGPT Pro plan, while the creator says a Claude Code session that long would likely hit limits within a few hours.
- Both models built a playable Nuketown-style first-person shooter clone in about 47 minutes each, from a prompt that told the agent to research how real shooters work before writing code.
- Neither one-shot FPS build got basic mechanics right — reload behavior was broken and neither included a working scope or zoom.
- AI-generated period music has crossed a real quality bar: ElevenLabs tracks styled as early-2000s songs were recognizable as pop-punk and 80s-synth pastiches, complete with sung lyrics.
- GPT-6 Astra's environment art consistently beat Fable 5.1's across all three builds, but Fable added small tactile touches Astra skipped, like an audible scroll-wheel click on the iPod.
- Fable's iPod build wandered past the brief on its own, auto-adding a 'Projects' and 'Now Reading' menu that pulled in the builder's actual GitHub repos.
- The creator's verdict after three head-to-head builds: GPT-6 Astra has overtaken Claude as his pick for the best coding model, for the first time in this comparison series.
What an 11-hour AI agent run actually buys you
WHAT TO LEARN
Two frontier coding models got the identical one-shot prompt on three builds, and the gap came down to patience as much as intelligence: the model willing to run unsupervised for hours produced the more detailed, more accurate result every time.
01Cold open: the rules of the head-to-head
- Every build follows the same rule set: identical prompt to both models, a single one-shot output, no revisions, and a live deployment at the end, so the model is the only variable.
- Judge any AI model comparison the same way: control for prompt, revision count, and deployment target before comparing quality.
02Build 1 setup: the retro iPod app
- Write the actual spec into a separate prompt.md file rather than typing it inline — it keeps the ask reproducible and lets you show or reuse the exact wording later.
- A build spec that asks for a real deliverable (working AI-generated music, not just a UI clone) tests more of a model's range than a plain UI request.
03Build 1 reveal: Fable's iPod
- Fable 5.1 took roughly 80 minutes on the iPod build and shipped tactile extras (scroll-wheel clicks, dents, scan lines) a plain feature list wouldn't have called for.
- An agent that wanders past the brief on its own (Fable auto-added a Projects/Now Reading menu pulling in real GitHub repos) is a real risk worth watching for, not just a quirk.
04Build 1 reveal: Astra's iPod + cost comparison
- Astra finished the same iPod build in about 37 minutes with sharper rendering, at roughly a third of Fable's cost, $13 versus $41, for a zero-revision build.
- AI-generated period music has crossed a real quality bar: multiple ElevenLabs tracks were recognizable genre pastiches, complete with sung lyrics, not just ambient loops.
05Build 2 setup: the D-Day reconstruction
- A prompt drafted by a separate research agent can look thorough but silently balloon the eventual build cost through extra input tokens before a single line of code runs.
- Citing a reference build (the Waterloo recreation) inside the prompt gives the model a concrete quality bar instead of a vague 'make it realistic' instruction.
06Build 2 reveal: Fable's D-Day
- A feature-rich interface doesn't substitute for the core visual ask — Fable's D-Day build had a dense, well-organized UI but thin 3D detail exactly where realism mattered most.
07Build 2 reveal: Astra's D-Day
- When a generated result looks subtly wrong (an unrealistic tide pattern), a model that stops to fetch and cross-check real source data produces a materially better outcome than one that doesn't.
- An 11-hour, fully unsupervised agent run didn't trip any usage limit on a $200/month plan, an important data point before assuming a long build will get cut off.
- More build time bought visibly better environment art (modeled rubble, a walkable village, detailed troops), while both models still hit accessibility issues like being unable to zoom the camera on demand.
08Build 3 setup + Fable's Nuketown FPS
- Telling an agent to research how a real system works before building a clone of it (real FPS mechanics before a Call of Duty clone) is a cheap, reusable step for any simulation-style build.
- Even a well-received build can ship with broken basics, both FPS clones had broken reload and no scope/zoom, so a one-shot build still needs a pass focused on core mechanics before it's called done.
09Astra's Nuketown FPS + final verdict
- Better character and environment art doesn't automatically mean better gameplay feel; judge a build on both axes separately before declaring a winner.
- When the same tool wins repeatedly on quality but loses on turnaround time, factor in whether you can actually leave it running unattended for that long before picking it for a real project.
Terms worth knowing.
- One-shot build
- A single prompt given to an AI coding agent with no follow-up corrections or revisions allowed, used here to compare raw model capability without human hand-holding.
- Fable 5.1
- The Claude model version used through Claude Code in this video, referred to throughout by its internal code name.
- GPT-6 Astra
- OpenAI's flagship model, run here through the Codex CLI and referred to throughout by its internal code name.
- Codex
- OpenAI's command-line coding agent, used here as the counterpart to Claude Code for running GPT-6 Astra.
Things they pointed at.
01:38channelPietro Schirano's original AI-built iPod (Magic Path)
Lines you could clip.
25:02
“I think Astra is the best coding model. For the first time, OpenAI has surpassed Anthropic.”
11:47
“If this was Fable, this would have hit limits in a couple hours at the most.”
19:32
“There's nothing I'd rather be doing than watching Astra revise each individual blade of grass in a D-Day reconstruction. This, I think, is what they mean when they say AI psychosis.”
07:27
“Thirteen dollars for that build, that is insane compared to Fable's forty dollars.”
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.
metaphoranalogystory
So I just give the exact same prompt to Claude Fable 5 .1 and OpenAI's brand new flagship model GPT -6 Astra across three builds. A working 3D iPod loaded with fake songs made to sound like it's 2003. A 3D recreation of D -Day you can scrub through.
And finally, a playable Call of Duty match on Nuketown. We've got the same rules for both models, a single one -shot output, no revisions, deployed live to the internet. The head -to -heads are back, baby.
This is GPT -6 Astra versus Fable 5 .1. Let's not waste any more time. Let's get right into it with these.
builds. We're going to kick it off with both these models. And here is the first prompt.
And the actual prompt here is very easy because we were doing slash goal and just having the models follow the prompt and the prompt .md, which I'll show you in a second. So I'm going to say the same thing to codecs as well. Slash goal, follow the prompt and prompt .md.
Do not stop till everything's complete. So we'll send these models on their merry way and take a look at this prompt .md. And here's our first build.
This is the prompt. So we are building a native Mac OS application that is going to be a classic iPod that's loaded up with songs that have been generated with the 11 Labs API that sound like they're from the early 2000s.
So this is gonna be a fun one. You can see this is a pretty detailed prompt. I will include it in the link below along with the generations.
And by the way, I got this from Pietro, Pietro Sciarano. He's from Magic Path. I mean, just look at this generation.
This is what he did with Astrid the other day. Pretty freaking cool. So I want to do something similar here.
What he's doing is he has his codex threads, but I want to actually include some music, but. Pretty awesome. So we'll see how close Astra and Fable can get to this and how they do with generating music.
Just checking in on the models. Been about 15 minutes. They're still humming.
Here's a little sneak peek. Little image. Oh, I don't want to spoil anything, but that's pretty cool.
Oh, look at this. Look at this thing. Just popped up out of nowhere.
What the heck? Wow. Okay.
And it's just using, I guess it's using computer use. I don't know what this is. I guess there's like a mirror of computer use.
Pretty cool. Fable, what are you doing? All right.
I don't know if you can hear this, but it's playing music right now. So it's very early 2000s. I don't mean to spoil this, but this is pretty awesome.
All right. Astra, surprisingly beat Fable 5. Looks like Fable is still creating the iPod.
Astro finished that in 37 minutes. So that preview is looking pretty good, but we will wait on Fable to see. All right, it's been an hour and 12 minutes and Fable is still working through this.
Oh, it looks like final render succeeded. There we go. Very close.
Meanwhile, Codex, oh, got a little preview from Fable. Don't want to spoil it. I don't know how to move this thing though.
ipod it looks like it's crashing okay finally fable is done with the build i know it says one minute ago i don't know what the actual time was it was probably an hour and 20 minutes looked like it was very thorough with its qa of everything but i'm just gonna say open the app so i can try it and we will see okay so here is fable's version pretty good pretty good i mean if we're getting picky what are these dents what did somebody drop this ipad but i mean ui Very clean.
We've even got some like scan lines here. You can see detail. So detailed that is.
And we've got like some dents here too. And then I can actually move this around. Oh yeah.
Okay. I can, that kind of rotates a little bit. Okay.
And then real question is, can we scroll on the scroll wheel? Oh, look at that. So here you can hear the little ticks too.
That's a nice touch. Okay. So I meant to just have music, but it looks like it added, I don't know what it added projects.
Oh wow. It has my GitHub repos in here. Okay.
How do we go back? Menu? Ah, look at that.
Okay, pinned. Now reading. Nothing open yet.
That's pretty cool. Shuffle tasks. Okay, say okay.
Settings. Brightness, clicker, sync now. That's pretty cool.
Threads. Okay, this just looks like a bunch of projects. All right, I did not mean to do that.
All I wanted was music, but that's a nice little test. Okay. then we have music albums so this is the music that it generated with the 11 labs api codex sessions volume one i realized i accidentally called this codex ipod so that's why it's calling it codex sessions here okay and here are five songs that have been generated with field and labs api let's check it out okay not really not really burly 2000s but okay next um maybe john mara kind of thing Oh, here we go.
Here we go. Pop punk. Is that the lyrics?
That sounds like good Charlotte. Okay, that's pretty crazy. What else we got?
Fluorescent. What is this one? Okay, that's like 80s synth.
Here we go. More pop punk.
Blink -182.
Oh, come on. That is crazy.
That's like Blink -182, maybe a little yellow card combination. That is insane. I haven't done these 11 Labs music generations in quite a few months.
These are getting freaking good. That's kind of scary. And let's ask Astra.
Please open the app. All right, you ready for Astros? Here we go.
I don't know how to drag this thing over to my other monitor. How do I drag this? No, I need to move it to another screen.
When I drag it, it just like rotates around. All right, hold please. This doesn't count.
It's already been built. This doesn't count as a second prompt. Let the record show.
All right, we got it working. There we go. There we go.
Now we can drag it. OK, I can't rotate it, though. It's really cool to do the rotation.
But OK, while it does that, check this out. Let's look at the difference. Open up again.
Look at this. I mean, you can just see the detail. way more detail.
I mean, it's the little stuff that Astro just excels in. This is more readable. The dial looks more realistic.
Like this is not, maybe this is like the first iPod, but this is like the, you know, the iPod video looks way more realistic. Text, like I said, and then when it gets this rotating, okay. One thing that Fable had was the sound effects, the little clicker.
Codex didn't add those. It's just that annoying. Whoa, where'd it go?
It's back. We still can't rotate it. It was really cool to see the sheen on the back, but.
Yeah, the clicks. It's a really nice touch from Fable, but let's see. I mean, everything else, I mean, this is a clear winner, but let's just go to the music that it made.
Okay, yeah, a little, like, Pearl Jam? Yeah. Oh, yeah.
We have lyrics. Keeps going on a loop. No lyrics, dang, for that.
Fast forward to lyrics. Good song choice, though. Okay, next one.
Why are they doing these lo -fi? That's not right. Terrible.
2004 Dance. Maybe I could buy it. Polaroid Weather.
Okay, I could buy this.
Bummer, we don't have lyrics. That would have been a nice touch. Weird lo -fi.
All right, we've got a clear winner. And then I think it added all the other things too. Yeah, projects, nice pinned feature, shuffle tasks, now reading, et cetera.
Pretty cool. Okay, I mean, just look again, so much better. And just in case you're curious too, here is usage.
for fable for this build so it looks like at least according to cloud code here this session forty one dollars cost forty one dollars available which actually is not that expensive for that long of a run i can never trust this in the cloud code desktop app though so take it with a grain of salt and then on the astra side doesn't actually show it can you tell me what the api cost would have been for this built inputs output tokens okay according to codex slash astra here about thirteen dollars for the original build here are the input tokens and output tokens including fixes i don't know where it got there was no fixes to be clear no fixes how do i drag this please open the app and that was it so i don't know what to talk about with fixes but okay and here are the rest of these numbers just because you're curious but it looks like thirteen dollars for that build that is insane compared to fables forty dollars both very cheap relatively speaking i mean usually fables in the hundreds for if you ask it what the weather is so that's that's pretty interesting all right next build we're gonna do the same thing slash goal i'm gonna say the exact same thing follow the prompt the prompt .md file in this folder do not stop until everything is complete give that to fable as well
bypass permissions. And for this build, if we look at the prompt at MD here, all of these prompts and all the live links will be available in the link below too, by the way. But you can see this one is rather extensive, just like the iPod builds.
Maybe too extensive. I get a little too, or not me. I had an agent do all of this and write all of this.
Got a little bit too in the weeds with this whole like eight kilometer coast of Viroville. In the west too, in the east. dog red easy green it just went and did a bunch of history research which i kind of told it probably didn't need to but this is just going to make this run expensive because there's gonna be a lot of input tokens and stuff like that but uh you can see all this and then i say like the whole point of this is to make it as realistic as possible and have some sort of scrubble timeline where i can go through minute by minute or hour by hour the d -day battle and really the whole goal is i want to see 3d generation that's super super realistic so that's what we're really putting the tests today and by the way i got this idea from dan where he had Astra recreate Waterloo.
And I just thought it was so cool. You can scrub through, same kind of idea. That's what I wanted to do D -Day.
But look how cool this is. I mean, it's just amazing. So that's what I want, like really realistic soldiers.
All right, 18 minutes in on the D -Day recreation. Looks like it's pulling up a bunch of terrain here. Bunkers, print.
Bluff contours. Measured terrain is now processed for Astra. They're doing very similar stuff.
All right. 43 minutes in for each of these models. They're still running through this.
This is a lot of work. I know. I'll write the unit movement data German strong points file.
Pretty impressive for both of these. Officially cross an hour. Looks like Codex is a little bit farther along.
It's reviewing a preview right now. Fable seems to still be humming. All right.
It looks like Fable has completed a little preview. preview there. Don't want to ruin it too much, but Fable is done.
It got done in about an hour and 10 minutes. Astra is still running, hour 13. Looks like it's QAing the tide.
The reverse of the wartime Omaha salt map includes tidal curves and beach profiles. I'm checking the scans, dates, units, and time convention to see whether it can replace the current half cosine model. Is it getting like the tidal patterns that day?
That would be insane. Yeah, confirming Astra was indeed doing this. It went and found the tides from June 6th.
tidal patterns because it felt like the tides that it was generating was not realistic enough. I mean, this attention to detail is just insane. And we've officially crossed the two hour mark with Astra.
It is now viewing images of this like a blurry image from I don't even know where it got this. And it looks like this is where the boats landed because it really desperately wants to be historically accurate. This is insane.
Look what it's saying here. The building symbols are verified, but the registration against the modern bluff is too uncertain for a convincing reconstruction. I'm retaining the annotated source image and documenting that gap.
The tested terrain and battle model remain the deliverable under review. Geez. I want to bring in some World War II history buff and see how accurate it actually got it.
This is just getting crazy. Look at what Astra's doing. I'm continuing with the visual gap.
The land still reads too much, like a photograph draped over a mesh. I'm checking the ve - vegetation scale and lighting so the bluff has visible death while preserving the measured terrain and playback performance.
I found why the grass looks sparse. The nearby patch uses individual blades spread over a large area. It is 11pm on a Friday too, by the way.
And you know, I just gotta say, there's nothing I'd rather be doing than watching Ashera revise each individual blade of grass in a D -Day reconstruction. This, I think, is what they mean when they say . Alright, so get this.
Ashera was not finishing this anytime soon. I had to go to bed.
So I came back the next morning and Astra finally completed this D -Day reconstruction in 10 hours and 51 minutes. Unbelievable. Meanwhile, Fable was done in 40 minutes.
Surprisingly too, I didn't hit any usage limits for this 11 hour build. Granted, I'm on the $200 a month Chachapati Pro, the 20X plan. But still, if this was Fable, this would have hit limits in a couple hours at the most.
I feel like we have to start with Fable first. I'm just having it deploy. please deploy to or so okay we've got d -day live on fable let's check it out Omaha Beach okay here it is all right first impressions pretty busy we've got these labels in the right we've got a timeline that's pretty cool so I can click different time stamps it goes to the different troops and squadrons this is like literally minute by minute that's pretty cool just very busy though a lot of information we've got sources with different labels and different colors we've got okay if i go overview okay that's nice and then we've got the tide here which is again very busy we have this timeline here that we can scrub through and we have the troops but it doesn't look like i can zoom in on them i can click them i don't think they're there's any 3d i wanted the actual troops looks like i can kind of see a boat yeah okay so it's it's a boat but it's A box.
Not what I said. I said I want it to be as realistic as possible. Now, granted, the rest of this is pretty cool.
We've got some nice detail in this landscape and the sand. Even when you zoom into the water, it's not great. Like, I've seen Fable do...
way better 3d work nice little light there though but i mean yeah this could have been done way better and then okay so we have like destroyer click it it's kind of weird and then destroyer like it's just got kind of lazy with the 3d work again i've seen it do incredible 3d so should have done a better job at this what else and then i'm just trying to click into different squadrons i can't zoom here or do anything can't zoom i can't zoom in any of these okay there we go and then there's just nothing okay so if we play this I'm just going to play.
This is really cool. So it does go through a minute by minute and I can scrub too. So we have, I'm sure this movement is accurate too, which is just incredible.
You can see them just advancing. But again, would have loved some troops there. Just got kind of lazy.
What else? Okay. You can click follow.
That's pretty cool. And then so you can like scrub and you could follow, except it doesn't really follow it. Go to first event.
All right. That's cool. I mean, if you're into war strategy or just history nerd.
you love this like my dad would love this they're advancing they're advancing and so we have like okay we have labels we have east bluff we have different locations that's pretty cool offshore overview clean how do i get the labels back i need to refresh okay overview overview is pretty cool and then we can go here and you can literally just scrub through the timeline also if you're a teacher this would make for a great teaching tool then we're just scrub moving through moving through you can see him and saying you can see him advancing destroyer i'm sure these like placements are accurate to like all the different squadrons and how they advance and the times and all of that.
And it keeps going and then okay, and then it just kind of ends. Okay, let's just click around a little bit more. So why here?
I don't know what that means. Plan versus actual. Oh, interesting.
Wow, it even got like the the original plan. And then what have they had to improvise? That's crazy.
And then we have all of these different squadrons here. Go back to plan versus actual. Why here?
Timeline strengths and losses. It's a little hard to read with these labels over the text. I mean, this is impressive.
This would have blown me away a couple months ago. But here we are. I'm not really that impressed, relatively speaking.
But let's see how Astra did on this. And it's 11 hour build. And if it really is that much of a difference.
Okay, here we go. Big reveal with Astra. D -Day reconstruction.
Okay, first impression. It does look somewhat similar. It's a little less busy.
Looks like the 3D. Okay, yeah, we've got some towns here. Oh, look, we got the town there.
That's pretty cool.
Okay, kind of fun. Oh, look at this. Oh, wow.
All right. Whoa. Jesus.
Okay. It looked a little weird when it zoomed out, but then when it came back in, wow, what the heck? It got like the rubble of the house.
Where did it even get this? Jeez. That is insane.
I mean, it doesn't look like a town, but it at least got the details of a house and it made it look like it's... You know, something that you'd see in World War Two in Europe. That is insane.
OK, then we've got the terrain here, the. Explore village so I can actually go in the village. So cool.
Wow. I mean, I don't know where the Germans are, but I wish I could just like kind of free float around a bit more, I guess, overview. OK.
All right. So we have we have all the troops here. Now, the question is, if we zoom in on this.
this again, it's, it's just like fables a little hard to control. I would like just to zoom in, but okay, look at these. Okay.
These boats are much more realistic, even though they don't really look like warships, but, uh, at least there's, there's some detail in the boats, but I mean, the, the actual detail, again, hard to navigate the actual detail and 3d work way better. I wouldn't, I wouldn't say way better with the 3d work. I mean, the fact that got the town like that is really impressive.
The, the waves. Again, trying to zoom in properly on the waves.
Would have been nice to navigate. The waves may be a little bit better. I literally can't scroll into the waves.
Okay, let's just go. Okay, inspect so we can go to different units. Oh my God, what the heck?
Okay, first of all, maybe not historically accurate. This looks like present day troops, but these are troops. And okay, they're like, all right, would have been nice to have like a, you know.
where they have six people i i didn't realize that the ar -15 was invented in the 1940s but here we are so okay do these troops move okay they're staying are they gonna move oh oh something happened oh they just moved okay so i just play this in real time okay kind of glitchy but i mean it got the troops again like let's look at these troops it's so detailed i wonder where it's getting this i wonder if there's like a public library or something that's pulling this from so okay we have the we have that one troop gap team seven and then we have the other troops here so if we go inspect troops okay more troops overview let's actually just play this now so if we go okay ground exploration okay some nice nice detail in the actual ground okay this is cool so you can play it from these different angles which is really really cool too but let's get it into a an overview and then we just play this and let's just scrub through this and see what happens okay so we see we see some troops this is already just way better than Fable should have been historically accurate but yeah we actually have the troops going and then we have some other ones coming in somewhere so dog red we have the boats here okay yeah the smoke's starting to form
That's a nice touch.
I'll let you play around with this, too, to navigate and see the actual troops. But there's a squadron there. I don't know if it's called squadron, too.
Military terminology is not very good. But I saw a boat. Okay, how do I get to the boat?
Got some smokes forming. We've got the USS. karmic how do i get to the uss karmic okay i want to see what this boat okay better boat still could have been more detailed but did a pretty good job i mean the waves are definitely more detailed look at the reflection of the waves okay let's just play this now and just watch this play out okay the us uss karmic is moving we've got troops moving over here we can go to the different troops you can learn more about these troops the actual navigation and the explanation is much easier to follow much easier to see i like the organization there uh again again i'll let you play around with this be the judge but this is this is we've got another clear winner all right in final generation we have a call of duty clone so i'm gonna do the same prompt to both these models just scroll follow the prompt the prompt .md do not stop until everything is complete kick it off with each of these models and i'll show you the prompt and here's the prompt build the best first person shooter you can possibly make in the browser get it perfect the map is nuketown small black ops map set in a nuclear test site suburb research this thoroughly go study how real first person shooters actually work i just had another agent just name off a bunch of things for it to go research to get this right we want to get as close to a call of duty clone as we can
possibly get it's really pushing the models to do that and then opponents free for all gunplay deploy it live etc all right so the call of duty clone got done in no time each of these models took about 47 minutes by the way too i don't think i've mentioned this yet Both models, effort level is high. Let's start with Fable again.
Copy this link, go here. Okay, Nuketown. Nice, good customization options here.
We've got our different weapons to choose from. Would have been nice to actually see an image of those weapons. You can adjust the sensitivity.
You can adjust the bot skill too. Volume, let's turn down the volume a little bit. Okay, deploy.
Let's do it. Okay. This already looks better than my Fable 5 Call of Duty generation.
Kind of glitchy here. little kind of boxy. Some stuff has a lot of detail.
Like, look at the lighting there, and then look at this car. It doesn't look very detailed. Oh, that's cool.
Okay.
This kind of looks like Nuketown, too, from Call of Duty. Okay. Jeez.
Tough. Not easy. What's going on with this glitching here?
It's going to give me a seizure. Alright, where's the enemy?
Jeez. Hard. I mean, the actual, like, shooting is pretty cool.
The, like, damage you're taking is pretty cool. Hard to aim. The people are pretty realistic.
It's more realistic than Fable 5, my generation from that. Let's make the bot skill just a little bit easier to compete with here. Jeez.
Computer's so good at this. Okay. Oh.
Okay, I'm using shotgun now. Jeez. Come on.
Alright, I'm gonna use assault rifle. One more, I just want to give you an idea of actually shooting at enemies.
Jeez. Oh my god, I'm terrible at this. Okay.
One more time.
So you can see just the enemy shooting. It's a little weird. Whoa.
Yeah, very weird. Jeez. The computer's really good and I'm terrible at this.
Okay, last one. I just want to show the actual graphics of the enemy. Jeez.
Because you can kind of see, you can see it here. You can see it here. Okay.
Yeah. Jeez. Terrible.
All right. So that's, that's Fable. Definitely better than Fable 5.
5 .1 was clearly a step up, but let's check out Astra's generation now. Copy this link.
Astra and just GPT in general, they just love this like forest green. The forest green is just everywhere. And it's like cream color.
It's the new AI dead giveaway. Okay. So we have our weapons.
Nice little icon here. preview of the map we got difficulty recruit and regular controls okay back to range okay i'm just gonna go assault rifle deploy to newtown okay a little bit better in the actual environment the actual map this one looks even more like newtown even has the sign there oh oh look at these look at that graphics look at these little players look how realistic these players are look at these guys That's crazy.
Shooting could be a little bit better. Oh, nice. Okay, I'm like crouched.
How do I reload? Would have been nice to scope in. That would be an easy add.
Oh, can you hit these guys? Like you can see just the graphics of the players way better. Look at these guys.
Okay, how do I reload?
Oh, whoa. Can I scope in? Nah.
Would have been a really nice touch to have the scope. I'm gonna have that next time.
But yeah, the 3D, the graphics are better. The gameplay plays a little bit better. Okay, I can't...
Yeah, the reload, that's a little weird.
Boom, gotcha.
Yeah, this is... I mean, this is freaking impressive. I love the graphics.
It really... I don't know where it got these characters, but they're... super realistic and so what it could do too it just didn't make the environment look realistic it just made the players look realistic but it can take these same same kind of player generations and apply it to the environment if you really want to get crazy with it but i mean just like this is a map nuketown i mean it looks so much like it i haven't played call of duty in years but it looks so much like it and look at these guys look at the way they're walking freaking crazy so i could add things like you know scope in maybe make it a little bit more realistic and taking damage i also can't reload let's just do this one more time because i want to just show more gameplay here boom so very very cool like just Freaking impressive from Astra.
So there we have it. Three builds in a head -to -head between GBT6 Astra and Cloud Failable 5 .1. I'm putting not even close again on this thumbnail because I don't think it was.
Astra was more creative. It had more attention to detail. It was more thorough.
The outputs were better looking. Yes, it took longer overall, but it is such a workhorse. And if that D -Day reconstruction was any indication, you can put this thing on a massive build and just have it run for hours and hours.
The best part is too, you're not... those dreaded usage limits where you would with Fable in like a few hours. The thing ran for 11 hours and I didn't even hit my usage limits.
So my same opinion still stands from my last video. I think Astra is the best coding model. For the first time, OpenAI has surpassed Anthropic.
So it's going to be really interesting to see what Anthropic does now, given that they don't have the leverage of the best coding model. Anyway, curious to hear your thoughts in the comments. Appreciate you as always watching.
Let me know any other future build ideas, anything that you want to see built with Astra or test. it against fable and feel free to play around with all the builds again all the prompts all the live links are in that blog post below and i'll see in the next one
The Hook
The bait, then the rug-pull.
The premise is simple and expensive: hand GPT-6 Astra and Claude Fable 5.1 the exact same one-shot prompt, no revisions, and see who builds better software with nobody holding their hand. Three builds later, one model needed almost 11 hours and a $200-a-month plan to pull it off, and still came out ahead.
Named ideas worth stealing.
00:00model
One-Shot Model Bakeoff Protocol
- Same exact prompt to both models
- Single one-shot output, no revisions
- Deploy live to the internet
- Compare time, cost, and output quality
A simple, repeatable way to compare two coding agents that isolates the model as the only variable — same spec file, same rules, same deployment target.
Steal forAny AI model comparison video, or an internal bakeoff before committing to a coding agent for a real project.
How they asked for the click.
VERBAL ASK
25:16link
“All the builds, all the prompts, all the live links are in that blog post below.”
Soft, low-pressure close — no product pitch, just a pointer to a companion post with the reproducible prompts and live demo links.
MENTIONED ON CAMERA
Visual structure at a glance.
