The argument in one line.
Across three identical one-shot build tests, GPT-5.6's Soul agent consistently out-designed and out-executed Fable 5, despite costing roughly eight times more and taking hours longer to finish.
Read if. Skip if.
READ IF YOU ARE…
- You use agentic coding tools (Claude Code, Codex, Fable) for real one-shot builds and want evidence on which model actually ships better UI, not just benchmark scores.
- You're weighing whether a slower, pricier agent run is worth it for design quality on a build you can't babysit.
- You build clone-style products — mockup tools, drop sites, data-driven platforms — and want to see what one unsupervised agent pass can and can't pull off.
SKIP IF…
- You're looking for exact prompts and step-by-step setup rather than a side-by-side outcome comparison.
- You don't use agentic coding tools and have no stake in a Claude-vs-GPT model comparison.
The full version, fast.
A creator ran the exact same prompts through Fable 5 and GPT-5.6 (via OpenAI's Codex/Soul agent) across three one-shot builds: a pixel-for-pixel clone of the mockup tool shots.so, an invented MSCHF-style viral product drop, and a from-scratch interactive NYC history platform. In every test, the slower and far more expensive Soul agent produced more thorough, more interactive, better-designed output — extracting working animation controls Fable skipped, building a multi-step interactive drop instead of a static page, and pulling in more real data on the learning platform. Fable only caught up after a second round of explicit feedback. The trade-off: Soul took roughly 6-9x longer and cost up to 8x more per build, and both agents' built-in cost trackers under-reported actual usage.
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 – 00:34
01 · Intro
Cold open framing the test: GPT-5.6 versus Fable 5 on three identical one-shot prompts, no revisions, deployed live.
00:34 – 08:13
02 · Build 1: Cloning shots.so pixel-for-pixel
Both agents run the clone-app-pat-pro pipeline against the mockup tool shots.so. Fable finishes in 30 minutes; GPT-5.6 Soul takes nearly three hours but extracts far more of the real UI, including working layout presets and an animation timeline.
08:13 – 15:03
03 · Build 2: The MSCHF test — invent a viral drop
Each agent invents an original MSCHF-style product drop and builds its site. Soul ships a confusing but highly interactive 'Human Resources' captcha-factory concept; Fable ships a clearer but static 'free shirt paid in fees' page.
15:03 – 23:51
04 · Build 3: An NYC learning platform from a rough idea
Given only a loose one-paragraph idea, both agents build a click-through NYC history site pulling from Wikipedia. Both underdeliver on the requested interactivity; after one round of feedback, both rebuild it as an explorable 3D city.
23:51 – 25:45
05 · The cost breakdown
Per-build time and cost for each agent across all three tests, plus a caveat that built-in usage trackers in Claude Code and Codex undercount real session cost.
25:45 – 26:26
06 · Final thoughts
Verdict: GPT-5.6 Soul was more thorough, followed directions better, and designed better across all three tests — despite costing and taking substantially more to run.
Lines worth screenshotting.
- GPT-5.6's Soul agent took nearly three hours and cost about $93 to clone shots.so, while Fable 5 finished the same clone in 30 minutes for $12.67 — yet Soul's clone was judged the clear winner on completeness.
- The MSCHF-style drop test cost $20.40 on Fable 5 (15 minutes) versus a reported $5.05 on GPT-5.6 Soul (30 minutes) — the cheapest build of the three came from the same agent that was most expensive on the other two.
- The NYC learning platform run cost $95 on Fable 5 versus $11 on GPT-5.6 — agent cost per task swings wildly and doesn't track consistently with either the agent or the build type.
- A single-page product idea with a clear premise (a 'free t-shirt' funded entirely by itemized fees) read as less creative than a more disorienting, multi-step concept (a captcha-factory 'prove you're human' shift) even though the disorienting one was harder to fully understand.
- Both agents defaulted to a conventional click-through, page-based interface for an 'interactive learning platform' brief until given explicit follow-up feedback demanding a 3D, immersive environment — the first unprompted attempt undersold what both models could build.
- One round of concrete revision feedback turned two flat, page-based NYC prototypes into two real 3D explorable city maps within about 30 minutes each — the gap was in the initial interpretation of 'interactive,' not agent capability.
- Built-in cost/usage commands in both Claude Code and the ChatGPT/Codex app under-reported real session cost compared to a third-party CLI cost tool (npx ccusage) run against the same session logs.
- A named, reusable cloning pipeline (recon, extraction, design spec, architecture, build, QA, fix loop, polish, deploy) run step-by-step produced a more feature-complete pixel clone than an agent that finished in a fraction of the time.
- Both agents deployed live, working builds to the internet from a single prompt with zero revisions across all three tests — the differentiator wasn't functionality, it was design depth and thoroughness.
- Soul was slower on every single build in this test, running against the common finding that GPT-class agents are typically faster than Claude-class agents — the creator flagged this as a possible anomaly, not a settled pattern.
The pricier, slower AI agent built better software three times running.
WHAT TO LEARN
Across a pixel-perfect clone, an invented product drop, and a from-scratch platform, the agent that took the longest and cost the most consistently shipped more interactive, more complete builds — and one round of concrete feedback closed the interactivity gap faster than a better initial prompt would have.
02Build 1: Cloning shots.so pixel-for-pixel
- A pixel-for-pixel clone test still measures real skill: matching padding, corner radius, animation timelines, and export behavior is much harder than copying static layout.
- The agent that took three hours and cost roughly 8x more extracted working interactive controls — layout presets, an animation timeline, adjustable light — that the faster, cheaper agent skipped entirely.
- Following a written cloning pipeline (recon, extraction, design spec, architecture, build, QA, fix loop, polish, deploy) step by step produced a more feature-complete result than an agent that finished fast without visibly working each stage.
03Build 2: The MSCHF test — invent a viral drop
- Asking a model to invent a viral concept, not just execute a spec, is a much harder creativity test than cloning — the two agents produced very different levels of interactive polish, a full multi-step simulation versus a single static page.
- A confusing but committed concept read as more successful than a clearer, safer one, because it took more creative risk and built more interaction, even though it was harder to fully understand.
- Telling an agent not to build 'just a static page' only worked for one of the two models — the instruction alone didn't guarantee a multi-step, interactive execution.
04Build 3: An NYC learning platform from a rough idea
- With almost no creative direction, both agents defaulted to a conventional page-by-page browsing experience instead of the interactive, immersive environment that was explicitly requested.
- A single round of concrete feedback was enough to get both agents to rebuild the same rough idea as a real 3D explorable city map — the gap was in the first unprompted attempt, not in agent capability.
- Design quality and information completeness didn't come from the same agent in either version — one produced better visuals with less pulled-in content, the other more comprehensive data with a less appealing interface.
05The cost breakdown
- The costlier, slower run didn't correlate with a worse result — it was the version judged better across all three tests.
- Subscription-plan agent tools don't reliably surface accurate per-session cost — built-in /cost and /usage commands understated real usage compared to a separate CLI cost-tracking tool run against the same session.
- If you're tracking real spend on agentic coding work, pull-based API billing gives a cleaner number than a flat subscription plan, where per-task cost is currently hard to audit.
Terms worth knowing.
- One-shot build
- A single prompt sent to an AI coding agent with no follow-up revisions, used as the controlled test condition before a second 'two-shot' round with feedback.
- Clone-app-pat-pro pipeline
- A named, reusable process for cloning an existing app or site step by step: recon, extraction, design spec, architecture, build, QA, a fix loop, polish, then deploy.
- MSCHF drop
- A limited, deliberately outlandish product release style pioneered by the collective MSCHF, known for stunts like selling individually numbered pieces of a cut-up sculpture.
- Wikimedia Commons API
- A public interface for pulling text and media directly from Wikipedia/Wikimedia's underlying data, used here as the fact source for the AI-built NYC learning platform.
Things they pointed at.
00:44toolshots.so
01:06toolclone-app-pat-pro skill
08:50productMSCHF
15:35toolWikimedia Commons API
24:47toolnpx ccusage@latest
Lines you could clip.
00:40
“Same rules for both, a single one shot output, no revisions, deployed live to the Internet.”
03:40
“The final runtime, by way, for GPT five six, two hours and fifty one minutes.”
07:55
“Soul just absolutely crushed it.”
24:01
“Four five six, as mentioned, it took three hours and costs almost a $100 in usage costs.”
24:56
“I can't believe I'm saying this, but I really think Soul took the cake here.”
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.
metaphor
00:00Alright. Well, Fable five has had the frontier all to itself for nearly a month, and OpenAI finally got the chance to answer. Thank you, government.
00:06GPT five six is live, and it's supposedly better than Fable in a number of ways. At least that's what the benchmarks say, so today, we're putting that to the test by giving the exact same prompt to both GPT five six and Fable five across three massive builds. A pixel for pixel clone of a design app, a viral internet stunt, and finally, an entire learning platform.
00:23Same rules for both, a single one shot output, no revisions, deployed live to the Internet. This is GPT five six versus Fable five. Let's not waste any more time.
00:32Let's get right into the building. And here on the left, we have Fable five. We've got GPT five six on the right in their respective harnesses, and we're just gonna paste in this prompt and get started.
00:42I just realized I needed to update to the new Chat GPT app too, the new app where they have codecs and works. We've got that running now. We've got Soul high running.
00:50We've got Fable high effort running as well, and here's the prompt. So goal, clone shots dot s o pixel for pixel using the clone app pat pro skill. I'll explain all this in a second.
01:00Basically, I go on and say what it is, recon, extraction, design spec, architecture, build, QA, fixed loop, polish, and then finally deploy.
01:10So shots dot s o, pretty cool mock up generator. In my opinion, that makes the the most beautiful looking mock ups. There's a bunch of these mock up generators, but I love this one.
01:19There's all sorts of customization you can do, the different styles, video, all different types of dimensions, phone mock ups. It is pretty sophisticated.
01:28I've actually tried to clone this before with probably Opus four five, and it it was not good. Because it's so complex, even like the interactions here, I want this to be I want us to reach feature parity here. That's what I told the agents.
01:39And that was gonna take a lot of work, so this is gonna be a good test. And while that's building as well, let me explain what I mean by clone app, PatPro. I use this to clone linear.
01:48If you've seen that video, I'll link it here if you haven't. But this is a skill I built out with Claude over a number of just attempting to clone different app sessions, and this is the process that we've landed on. It's similar to the process I laid out to the agents in the prompt.
02:00It just gives more detail. I will leave this in the description for you to grab if you're looking to clone apps yourself. So Fable, at about twenty seven minutes, is already done.
02:09And then Soul looks like it's still on step one of eight. It looks like it's just viewing the images right now. So interestingly, Soul is actually slower than Fable, which goes against a lot lot of people's findings and the benchmarks as well.
02:20Usually, Fable is much slower. This could be an anomaly, but we'll just let them keep running here. Fable five has already finished.
02:26It finished in about thirty minutes. I'm not gonna look at this yet, but here's all the work it did. Seems pretty comprehensive.
02:32I'm a little concerned how quick this was. Soul is following the clone app skill much more closely, saying, I need to interact with it, say, begin extraction to resume the goal.
02:43So it hasn't even started extraction yet. Begin extraction. So that whole time, I think, was just taking screenshots.
02:49Then it's gonna look at the code in shots at s o. Like, it hasn't even started architecting this yet. 48 later, it has started the design spec.
02:56We'll stop before architecture. Soul is being incredibly thorough here. Let's see if it pays off.
03:00Oh, boy. It has been an hour and forty one minutes. Okay.
03:04I I need to say continue now. Continue. And it is on the build.
03:08It looks like it completed the build, and it's proceeding to QA cycle one. It has officially taken over an hour more for five six Soul. Alright.
03:17QA cycle one has been completed. The agent is starting polished now. We are approaching the two hour mark with GPT five six Soul.
03:24This is insane. It's finally done, but then it now wants to QA. So QA'd it.
03:29It's worked locally, and then now it wants to QA with a live link. But we are officially at two hours and forty five minutes with GPT five six Soul. This thing is, uh, is now live.
03:40So I feel like we have to view Fables first. The final runtime, by way, for GPT five six, two hours and fifty one minutes. But let's see how these look in the browser.
03:50Here's Fable five's version. Um, okay. K.
03:54I can change a window. There's nothing on the left here.
03:58Let's just show shots at s o again. Here is shots at s o. So you can see you can add media.
04:03There's all of these different styles on the left. Core Fable definitely did not add that.
04:09We have some gradients, other gradients. I mean, this isn't bad. Okay.
04:13We yeah. We've got a nice little window UI. Phone.
04:16Okay. That doesn't look like a phone at all. Oh, we can go tall, nine by 16.
04:20Okay. No. We wanted like the true phone mock ups, which don't appear to be showing up.
04:26We have all this like padding and whatnot that doesn't appear to be working. Oh, yeah. Okay.
04:31There's a little bit of radius here, so that changes padding. I don't think that changes.
04:36Three d tilts. Okay. That works.
04:37Yes. That's pretty good. Remove.
04:40Okay. Drag or paste. Screenshot actually works.
04:43Let's see. Oh, cool. Oh, that looks good.
04:46The little phone except it's not truly a phone. LinkedIn bullshot.
04:50Yeah. You can do a background image. You can do a gradient.
04:53You can do transparent, plain. So not bad from Fable, but honestly a little underwhelming.
04:59Let's see how Five six does. Copy this link. Go here.
05:06Woah. Okay. Um, dang.
05:09This is not shots. S o. This is GPD56.
05:13This is shots. S o. It looks nearly identical.
05:16Wow. It even has this, like, magic preset. I don't know if that yeah.
05:19That doesn't work. Okay. I mean okay.
05:22Yeah. I mean, it got it got some of these images. So so it actually able to export, if we go here, some these images.
05:28I mean, obviously, the formatting isn't right. This looks super weird. But like, does this work?
05:33Let's see. Let's do not an iPhone, something else. My pad pro.
05:38Wow. I think that might work. I think that's actually working.
05:41MacBook? Didn't get the MacBook. We go here.
05:44Yeah. You can see. It's a bummer because these images, I think, are, like, available in the code.
05:50So the agent should have been able to extract this. But, I mean, this is let's see if it works. Drag or paste images.
05:57Does this work? Oh, look at that. It works.
06:01And then style, that doesn't do anything. Okay.
06:03I have to go back to phone. Lavender. Alright.
06:06Yeah. Change the style. Display.
06:08Like, I don't know if this is a proper phone mock up. Yeah.
06:11No. No. Shots.so looks much more realistic.
06:14Shadow adaptive. Does that do anything? Don't think so.
06:18That does something. Adjust light. Does that work?
06:21Oh, wow. It does work. Jeez.
06:23Adjust light. The same thing in shots. S o.
06:26What else? Frame? Does this work?
06:29Okay. Frame. Let's see what it says in shots.
06:31S o. Portrait watermark b g effect portrait. What does it do in shots.
06:36S o portrait. It didn't get that. Well, it it this there's like just way too many options here in shots at us.
06:43How do I even work this? Lens blur with stage. Oh, cool.
06:47No. It got that. Portrait.
06:50Oh, no. Okay.
06:51It has shadows. Oh, lens blur. There it is.
06:53Stage? Doesn't really do anything, but it at least got those buttons and it actually works. Stage, you can adjust these.
07:02Shapes, add shapes. Background, you can change the background, change the color. You can upload images.
07:07Can Unsplash? Can you actually grab from unsplash? K.
07:11No. Don't think so. Magic, does this do anything?
07:13Okay. Something. Backgrounds work.
07:17Let's see if watermark works. Okay. I don't think watermark does anything.
07:22Oh, okay. Look at that. Got some text here.
07:24Can't really move the text. And then look at these layout presets. Look at this.
07:28These are actually working. A two by two. Wow.
07:31Desktop. Wow. This is just crazy.
07:33This is an insane difference. Again, look at Fable. Fable just all it did was give some canvas sizes and some different frames, and then five six just went ham bony.
07:45Animate? Does this animate? No way this actually works.
07:47I don't know how to work this animator. Add animation? Okay.
07:51That doesn't do anything. But the fact that it even has this little animator timeline is pretty darn impressive. Anyway, .
07:58Alright. So we have a clear winner, five six Soul. And I just went back and checked because I didn't know if I didn't set a goal with Fable five, but it was the exact same prompt via goal.
08:10So you can see just how much more thorough five six soul really is. So that's the first test cloning an application. But a clone really is just following directions.
08:17It's copying what's already there. The model never has to invent anything, however difficult it might be to clone an application like shots at s o. But next, we're gonna see how these models introduce their own creative thinking with what I'm calling the mischief test.
08:31So mischief, if you're not familiar, are the kings of viral Internet stunts. They've been doing this for years and each I don't know. It used to be every week.
08:39I don't know how often they do it now, but they have these drops. And let's just look at one here. King Solomon's baby.
08:44So they're so they're cutting up a $100,000 sculpture and selling each one of the pieces, which is just ridiculous.
08:51I mean, have so many of these ridiculous sons. And they're just incredibly outlandish with their ideas, which an AI is notoriously not great at. But we're gonna try and see how well these models do with out of the box thinking.
09:00So we're gonna have each model invent their own mischief drop and design a website. But we will go back to Fable and five six here, paste this prompt in. I'll put this one on screen as well.
09:13But what we're saying is invent an original Mischief style drop and build its full drop website. We'll give some information. We say study this, work out what makes it a Mischief drop land, come up with at least 10 ideas, then choose the best idea, and then build out a website.
09:26And the models are off and running. Hopefully, it won't take five six soul another three hours or I will be recording this into the wee hours of the morning. K.
09:32Both models have completed the mischief drop. Fable took about eighteen minutes to get this done. Five six, again, pretty slow.
09:38It's a goal of achieving fifteen minutes, but I think it took longer. Probably more like thirty. So let's look at five six soul first.
09:45Remember, we are grading not just on the website, but on the drop idea as a whole. Okay. Human resources.
09:52One human verified stamp. $0 plus 100 captchas. Not exactly clear what that means.
09:58Okay. Clock in. Oh, that was nice.
10:00Look at that. And then alright. So we have a shift here.
10:03Welcome candidate. We made something to prove you're irreplaceable. A robot made it.
10:07Oh, interesting. Okay. Human config, 99%, six robots, zero humans.
10:12Still trying to understand what this is. Select all squares containing work with dignity. Care?
10:17Union? Teach? Synergy?
10:19What what what do mean dignity? I'm still not. Alright.
10:22It just keeps going. The last job only humans can do is prove they're human. Every CAPTCHA is a tiny job interview conducted by a machine.
10:29Identify the bus, transcribe the word, rotate the animal, produce a labeled example. The object is molded, inked, assembled, inspected, and packed by machines.
10:37You do not buy the stamp. You train the factory that made it. Still a little confused.
10:40Okay. This is nice, though. Got this conveyor belt.
10:44Okay. So this design is pretty incredible. Policy, you're independent enough to be responsible and dependent and dependent enough to be unpaid.
10:53All successful work proves you're human. All failed work proves you're human. I I'm still just like I I I think I know what's going on.
10:59It's like a captcha to prove that you're human. K. And then we have, like, rejected drops.
11:03Okay. This is the idea. It's nice.
11:04Okay. We'll look at those in a second. But first okay.
11:06So just clock in. Nice animation. K.
11:09We're just gonna select six. Select all scores containing work with dignity. Okay.
11:14So it didn't say six. Selection will be used to automate future selections, accuracies, whatever management says it is. K.
11:22Shift completion. Oh, okay. Move the slider until compensation is fair.
11:26K. I don't I don't know what's going on. The slider maybe doesn't work.
11:29Okay. No. Never mind.
11:30Fairness adjusted for market conditions except compensation. Type the text to prove you understand the text. This is, a very mischief thing to do where you don't even really understand what they're doing.
11:41I consent to unpaid labor. Which face was generated by a human? What what is going on?
11:46Your label your labor has made you eligible to stop laboring. You produced 100 machine readable proofs of humanity. Claims I mean, all of these interactions that suffered killer.
11:55Okay. Shift complete, human verified, print temporary.
11:59Okay. Wow. I mean, some of these functions are really quite complex, at least for a one shot.
12:05Got this timer here. Again, maybe somebody watching this understands what's going on, but I'm a little lost. Still, it is pretty cool.
12:12Pretty cool. Pretty complex. Good web design.
12:16Good interactions. I love the aesthetic it leaned into. It's definitely outlandish, although it doesn't make a whole lot of sense.
12:22But there we go. Okay. So, like, let's look at your rejected initiatives.
12:25Zero star hotel, book a nonexistent hotel room so you can leave a verified review place you've never visited. 1,000 identical objects each packed until it's the last one.
12:35Okay. I'm not gonna read all of these. Alright.
12:37Like I said, we'll have the link for you to interact with us, but that was pretty good. That was that was creative enough. It was out there enough.
12:44The web design looked looked beautiful. Some nice little interactions. So that's five six soul.
12:48Now let's look at how Fatal did. The drop is live. Free shirt.
12:53Oversell.app. Okay. Similar kind of idea.
12:57We've got that carousel at the top. The free shirt. The first genuinely free t shirt in the history of commerce.
13:04Okay. Kinda lame, but add a tip for checkout experience. Your tip has been billed as a processing fee.
13:09Okay. Get your free shirt. Why is it free?
13:11Fees collected to date. Price of the shirt. The checkout is the product.
13:16Other companies hide their fees at the end of checkout. We're not other companies. Our checkout is the drop.
13:19The shirt costs nothing. Okay. So I'm not charging fees, I guess?
13:23Every fee on your receipt is fully itemized. Click any fee to see what it's made of. Fees are made of smaller fees.
13:28You may remove any fee at any time. Removing a fee incurs a fee removal. Okay.
13:32It's just like alright. That's kinda being, like, funny with fees. Okay.
13:36Alright. So you remove and then it adds more fees. Okay.
13:40Interesting. Okay. I I mean, nice little, like, receipt interactions.
13:44The idea is understandable, but it seems dumber and not as creative, to be honest. You've never bought a product, so you've only ever paid fees. The product is a loss leader for the fee.
13:54Not entirely true, but alright. I guess they're yeah. Sure.
13:58Concert tickets, that applies. Is this shirt really free? And why do I pay $38.47?
14:03Okay. Yeah. I mean, this is like it makes sense, but that's not exactly out there.
14:08The design looks good. Why is it free? Can I click how this drop was chosen?
14:12Okay. There's there's the list of the other ones. I'm trying to click to see if there's anything else here.
14:17It's it's just one static page, which is a little disappointing even if the design is pretty good.
14:22Just adds fees. Okay. I see.
14:24But, I mean, again, like, look at all of the interactions and, like, things you click through with Soul's example. And, I mean, I think we've got a clear winner here again.
14:34I am really blown away with how well Soul is doing right now. Just in terms of designing the experience, it's just so much better. Yeah.
14:41The idea is a little confusing, but this is so much more interactive than a static page, which I told the models not really to do. So Fable should have caught that. But, I mean, with both of these tests, this is not even close.
14:51Now for the last test, because, yes, we did copying, we did inventing. With both of these builds, you were giving the models some kind of guidance. The last thing I wanna do is test how well they do with basically no direction at all.
15:03So I have this half baked idea around this New York City learning platform. I'm studying to be a New York City tour guide for no other reason than it's a good excuse to learn about New York City, and I love history. And it's just so boring to be reading Wikipedia all the time, so I wanted to do something.
15:17I wanted to build something that's super interactive. I basically want an ability to see different buildings, click into them, and get some kind of history. But not like boring history like this building was enacted in 1898 stuff, but more fun history facts.
15:31And I I don't even know what I mean by fun history facts, but we'll let the models make those decisions. So this is what the prompt is. Go to both of the models, paste this prompt, and and here's the prompt.
15:42I'll put it on screen now. But what I'm saying is build me an interactive learning platform for exploring New York City. Here's the rough idea.
15:48It's deliberately rough. Figuring out the rest is the test. I wanna explore the city, see the skyline, click into different neighborhoods, down to individual buildings and landmarks.
15:56Everywhere I click, I get genuinely interesting facts. I'm having the models pull from Wikipedia. Think it's the best source of information, so we're gonna use the Wikimedia Commons API.
16:05Every fact has to pass the I didn't know that test. And then finally, don't know how they're gonna build this out. I don't know if it's gonna be some sort of interactive world, if they're gonna get into three j s, some kind of three d interactive experience, but I'm leaving it up to the models to decide all of that.
16:18And I'll just give that same prompt to Codex and let these models get going. Alright. So for once, Fable is working more than five six Soul.
16:27Five six Soul is finished in eleven minutes. Fable is still humming. We're getting a preview here of what it's creating, but we'll let Fable finish up here.
16:35Alright. Here we go. So Fable is done now.
16:38Took about fifteen to twenty minutes. So let's start with Fable first, and okay.
16:44I just kinda built a website. I don't love this design. Oh, look at that.
16:48You can okay. Alright. Nah.
16:50Not bad. Not bad. Actually, click these.
16:52The buildings, like, that doesn't really look like the Chrysler Building. Come on. Empire State, 1 World Trade, Woolworth, Flatiron.
17:00Like, I don't think these are oh, I guess these are technically probably placed together in a Woolworths and Fideye. Yeah. Flatiron, Empire State yeah.
17:07So maybe it did. Okay. Kind of yeah.
17:09It's kinda doing the skyline. I just don't I don't like this design at all. I hate the colors in these cards.
17:15Very a I slop looking. But but gonna try to get past that and just click into some of these and see what happens. So click to uncover.
17:23K. Nice photo photo from Wikipedia. A b 25 bomber hit it, and an elevator operator felt 75 floors and lived.
17:31Wow. Okay. That's alright.
17:33That's a good fun fact. That's what I'm looking for. Okay.
17:35Then tap to uncover. The building has its own zip code. The Elegant Spire was originally yeah.
17:40For dirigibles. I knew that one. The little, uh, the little zeppelins.
17:44The idea was people thought that we would be traveling in blimps everywhere. Never really took off. But the original idea was it was like a docking station for these blimps.
17:51Uh, I'm not gonna turn this into New York City history fun fact hour, but, um, okay. So for for the first years, it was mocked as the empty state building you It opened during the depression.
18:01Uh, okay. Cool. Yeah.
18:01Yeah. On the map. Not bad.
18:03I I would have liked to, I don't know, keep this on the page. It feels like there's a little too much friction. We actually have to click in and review all of these.
18:10Would have been nice to just have it on one page. But some good fun facts. Alright.
18:14Brooklyn. Okay. So you can click into the boroughs.
18:16You can see this. Like, I don't really wanna do that. Okay.
18:19But I guess I can click into this. Alright. I mean, built out like, is quite a bit of data here and you click into these and then it has dedicated pages for all of this in the different neighborhoods.
18:27You can also keep going. So also in Greenpoint, reveal these. But again, like the actual user experience, how this is organized, I think could have done a much better job, not to mention the UI is trash.
18:37The power broker city. Alright. But, yeah, let's get in some Robert Moses fun facts.
18:41Yeah. So it's just built out this pretty extensive knowledge graph here from Wikipedia. Oh, nice.
18:46We can click those. See, this is what I would have liked. It's like some sort of map you click and it has like some sort of pop up.
18:51I mean, ideally, it's like in some sort of immersive environment. So this feels, very static. And I even told Fable, I want this to be interactive and immersive.
18:59Didn't really listen. I wanna keep clicking these because it is it is interesting, but it doesn't feel like a platform where you want to learn. Anyway, that is Fable.
19:08Let's check what five six did. So go here, explore city unscripted.
19:17Nice little g sap there. Oh, nice little animation. Okay.
19:21Much better design. How is five six just crushing Fable and design? Alright.
19:27Not not not great. Not great. Little blobs here, but okay.
19:31We got something. Choose a burrow. Alright.
19:33I mean, design is better, but still both of these, a little disappointing. I thought they would. Maybe I should have just said do some sort of three d here.
19:41But okay. Manhattan, three pass. Why is there only three neighborhoods?
19:45That's kinda weird. The island grew. Okay.
19:49Lower East Side. Alright. We we got some buildings.
19:52Alright. A little bit better. At least I'm, like, getting this visual even though it's, like, cropped in and I can't really tell what the building actually looks like.
19:59And then I can't go back? Like, why why do they choose these three neighborhoods? Should've should've built it out more.
20:05Harlem, Glance, Staten Island.
20:08Okay. Couple. Alright.
20:10Yeah. I mean okay. I mean, yeah, that's kind of it.
20:13Both a little disappointing. Soul, I definitely like the design better, but it was lacking information.
20:18Fable, just terrible design. Here's what I'm gonna do. Uh, we're gonna go back.
20:22This is gonna be a two shot for both of these. I'm gonna give them another chance. What I really wanna do is, like, stress the interactivity and the engaging experience and probably just push them on building some sort of three d environment and then just see what happens.
20:33Okay. Here is my feedback. I'm just gonna hit enter on this.
20:36Let these models run. But what I'm saying is this is not interactive enough. I said really engaging.
20:42Like, I'm just clicking from one page to another in a static site. That's not what I wanted. Too boring.
20:46Gets confusing. There's too much friction. We're having to click all these pages.
20:48It makes it hard to learn. It should be much more visual and immersive. And then I just said slash gold.
20:52Don't stop me until you've built a completely interactive and immersive learning environment that ties visuals to these facts. And I said the same thing to five six. The only thing that I added was, also, you didn't pull enough info from Wikipedia.
21:03Why only a few neighborhoods and landmarks? So, I mean, there's a good test too of how these models take feedback and revise. Models are done updating our three d demo.
21:10Let's look at Fables updates first. This took about thirty minutes. And okay.
21:14Enter the city. Okay. Woah.
21:17Alright. This map looks pretty good. That looks pretty accurate.
21:21Bridges are kind of all over the place. Okay. We got the spire on 1 World Trade.
21:25We've got an attempt at bridges. Oh, look at this. Okay.
21:29Can I Okay? Way better. Way better.
21:32Boom. Let me go with smaller face here. Yeah.
21:34Way better. Okay. WASD would have been a nice touch.
21:37I don't think I can move around. Yeah. Nice.
21:41Nice. Chrysler Building. K.
21:43Kinda hard to get. Nearest dark building, Grand Central Terminal. If you could somehow like mock this up in three d, it would be really cool to turn this into some sort of like VR experience.
21:52That would be killer. I'm gonna make that up. I feel like these models are good enough to build something like that out.
21:57Okay. I mean, hard to navigate. Can't really zoom.
22:00Scroll to zoom. Okay. Scrolling.
22:02Alright. Yeah. Drag to orbit.
22:04Okay. Okay. And then Alright.
22:07Yeah. Like, do we click? Okay.
22:09Some of these bridges, some of these are not pulling the correct images, but I mean, this is pretty darn good. So we have this, like, interactive map that we can orbit around.
22:17We can zoom in. Good job, Fable, on taking feedback. Oh, what do we got?
22:20A boat over here? Look at that. There's a boat just going around.
22:23Interesting. Okay. So now let's go to 56.
22:27See how well it did. City unscripted. Okay.
22:29Oh, nice little animation. Okay. Different kind of take on this.
22:34I liked Fable's map a little bit better. K.
22:37Drag to orbit. We have these nice little oh, nice. Okay.
22:41You got some of the images. Yeah. Trinity Church even though it doesn't look anything like Trinity Church.
22:46Okay. And I see you can just like go to these. Yeah.
22:49Alright. I I actually like this. I like this experience a little bit better.
22:53If it if it combined this with the map view from Fable, it would be really quite good. Also hard to scroll. I'm scrolling back and it just snaps back in.
23:02That needs to be fixed. Let's x out of this. I can't x out of this.
23:06Next beacon. Okay. Yeah.
23:08This is way better. Again, some like true to life three d, which you could actually get from like Google Earth where the Google Tiles API, which I did in my Unreal Engine MCP video.
23:19So you could probably combine that and create, a a true to life accurate looking three d model, and then just self select some interesting buildings. Financial District, Lower East Side, NYC.
23:31Okay. There we go. So you just have to click back.
23:32Alright. We've got, like, Prospect Heights. Alright.
23:36I mean, both of these have did a pretty good job. I think I'm gravitating more towards Fables generation, but I do like how you can control these on the bottom. The stuff on the left actually grabbed those images properly.
23:45They weren't a bunch of just broken images. The actual panning is a little funky, but you can see how well these models do when you really push them on feedback. So a quick wrap up first on costs.
23:55Here's what we're looking at. The shots dot s o clone took thirty minutes. For Fable, it costs $12.
24:01Four five six, as mentioned, it took three hours and costs almost a $100 in usage costs. Granted, this is on my subscription too. I'm not paying this in API cost, but this is what it would cost if you were paying through the API.
24:12And then the mischief drop took Fable five fifteen minutes, cost $20. Five six, it took thirty minutes and apparently cost $5. I'll give a caveat here in a second, but that's what I was looking at when I actually looked at the usage.
24:22And then the NYC platform took Fable five twenty minutes but cost $95. And five six, eleven minutes or a little bit longer, twenty something minutes including the revisions and it cost $11. Now granted, this is what I just did when I had an agent look at the logs.
24:38And even in it's a little confusing even in Claude code and in Codex. Like, I I can't find a way to actually see this in Codex. In Claude code, you can just go slash cost or slash usage.
24:47And for example, on the learning platform, it said $99 here. For the rest of these, it didn't seem accurate because I ran slash usage, it says cost $3. And then for the shots at s o, I ran usage as well.
24:59And you can see, it just doesn't even show up with the API cost. So it's kind of all over the place. There's still some bugs I think they're actually working out and displaying displaying this usage.
25:07But I think a more accurate way to do this is to just run this bash command here, which is just if you're curious, n p x c c usage at latest session. There's one similar for codecs, so I had an agent checking this on the side to get those actual costs. And those costs seem to be more accurate than what's actually displaying in the usage here.
25:22So take all those with a grain of salt. I'm working on having some kind of system that will track this more thoroughly just from the start next time. It's much easier to track this if you're in, you know, something like OpenRouter where you're just doing pure API costs.
25:34Think, I like I said, there's just some sort of bug or there's something going on where it's hard to tell what your session cost is when you're on these subscription plans. And on the $5.06 side, I'm on the the $100 a month plan. On Cloud Code, I'm on the $200 max plan.
25:45Anyway, that is going to do it. So I can't believe I'm saying this, but . It was more thorough.
25:52It followed directions better. Its designs were better. Its creative thinking was better.
25:57But I don't know. I'll leave it up to you to be the judge. I'm curious what you guys think.
25:59Let me know in the comments. I just honestly thought that for sure Fable would be much better at these kind of big, you know, creative builds. And if you wanna see how g p t five six stacks up against other models with a real breakdown of benchmarks, all of that, you can check out this video here.
26:12And send me your build ideas too. Fable five six, GBT five six Soul, whatever. I wanna keep stress testing each of these models, figure out, you know, where they're actually strong, where they fall apart, and just bring all of these findings back to you guys.
26:23So as always, appreciate you watching, and I'll see you in the next one.
The Hook
The bait, then the rug-pull.
OpenAI's GPT-5.6 finally had an answer to Fable 5's month-long run at the frontier — so the only way to settle it was to give both agents the exact same prompt, three times, and watch what they actually shipped.
Named ideas worth stealing.
01:06list
The clone-app-pat-pro pipeline
- Recon
- Extraction
- Design spec
- Architecture
- Build
- QA
- Fix loop
- Polish
- Deploy
A step-by-step process for cloning an existing app: capture the live product via screenshots/interaction, extract assets and structure, write a design spec, architect the build, implement, then run QA and polish loops before deploying.
Steal forany project cloning a competitor's UI or reverse-engineering a reference product
How they asked for the click.
VERBAL ASK
26:00next-video
“if you wanna see how GPT five six stacks up against other models with a real breakdown of benchmarks... you can check out this video here. And send me your build ideas too.”
soft, low-pressure — points to a benchmark video and invites viewer-submitted build ideas for future tests, no hard sell
Visual structure at a glance.
