$100M AI Companies Are Officially Cooked (Fable 5)
Four one-paragraph prompts on the highest reasoning setting turn into a real SaaS funnel, a working Minecraft clone, an open-world GTA-style game, and a full agency-grade ad campaign — all in one sitting.
A single one-paragraph prompt on the highest reasoning tier can now produce a finished, sellable deliverable — a funnel, a game engine, or an ad campaign — in one long unattended run, at a real dollar cost per job.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You run an agency or freelance practice and want to know whether AI can now produce billable-grade deliverables (funnels, ad campaigns, games) from a single brief.
You're deciding whether to spend premium AI credits on big, expensive one-shot jobs versus everyday small tasks.
You want a realistic read on cost, time, and failure modes of long-running autonomous coding/agent sessions, not a marketing demo.
You're evaluating multi-tool agent orchestration (calling image, video, and outreach tools from inside one model session) for production use.
SKIP IF…
You're looking for a step-by-step tutorial — this is a demonstration and review, not a how-to build guide.
You want unbiased benchmarks — the creator explicitly skips numbers in favor of subjective builds.
TL;DR
The full version, fast.
The creator tests Claude's top reasoning tier with four single-paragraph prompts and no follow-up guidance: a complete go-to-market funnel for a fake SaaS (real Next.js site, checkout, 11 generated images, 6 video ads, and cold outreach drafted via Clay and Gmail), a from-scratch Minecraft clone with its own chunk and lighting engine, a full GTA-style open-world game with physics, traffic AI, and a police wanted system, and an agency-grade brand launch campaign (research, 16 images, a cinematic ad film with voiceover and score, a pitch deck, and a live page) built inside a separate multi-agent tool. Each build ran for roughly one to two hours and consumed hundreds of thousands to close to a million tokens. The model hit rate limits multiple times, self-diagnosed its own bugs and fixed them, and clearly labeled which parts were AI-rendered versus its own reasoning. The conclusion: this tier is genuinely capable of one-shotting entire finished projects that previously took a team days, but it is expensive, slow, and meant for hard, high-value jobs, not everyday tasks.
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.
Creator skips benchmarks and previews four one-prompt builds run on the highest reasoning setting.
00:34 – 01:02
02 · What the model is and why it was pulled
Explains this is Anthropic's top model for long unattended jobs, temporarily taken offline over a government security concern, now back with tightened guardrails.
01:02 – 02:24
03 · Build 1: go-to-market machine (setup)
Prompt: build a complete go-to-market machine from scratch for a fictional missed-call-recovery SaaS called Ringback AI, using connected tools for images, video, outreach, and code.
02:24 – 03:52
04 · The funnel it built as a real Next.js site
A six-page real site: sales page, opt-in, Stripe test checkout, order bump, upsell, mocked dashboard with fabricated but self-aware testimonials, plus an 11-image generated asset set and a real 8-page lead magnet PDF.
03:52 – 05:44
05 · Real cold outreach through Clay & Gmail
It generated six video ads, then used Clay to find and verify 30 real business owners across five cities, writing a personalized cold email to each and saving them as Gmail drafts.
05:44 – 06:11
06 · No other model could do all that from one prompt
Creator names the rate limits, lost task IDs, and safe fallback to self-addressed drafts as honest rough edges, then puts a real time/token cost on the two-hour build and claims no other model matches it.
06:11 – 08:25
07 · Build 2: Minecraft — Opus vs Fable
Shows the prior model's simple voxel sandbox, then the same one-prompt task given to the new model: full custom chunk system, lighting engine, day-night cycle, and functioning hotbar.
08:25 – 09:05
08 · Not even a fair fight vs Opus
Direct side-by-side comparison; the build survived a rate-limit pause mid-project and resumed exactly where it left off, finishing in about ninety minutes.
09:05 – 10:29
09 · Build 3: GTA-style open world (setup)
Prompt: build a drivable, walkable open-world game with its own vehicle physics, procedurally generated city, and no copied assets.
10:29 – 13:08
10 · It's an actual game
A twelve-by-eight-block city with ~600 buildings, working car physics and drifting, AI traffic and pedestrians, a full day-night cycle, and a wanted system with police chases that the model stress-tested itself before delivery.
13:08 – 13:21
11 · Its review agents died — so it found 4 bugs by hand
A self-assigned five-agent code review team all hit a session rate limit; instead of hiding it, the model reported the failure and manually found and fixed four real bugs.
13:21 – 13:57
12 · Build 4: Kaya coffee campaign (setup)
Introduces a separate multi-agent orchestration tool and gives it one prompt to act as creative director for a fictional coffee brand's full launch campaign.
13:57 – 15:35
13 · Real research, 16 images & the cinematic ad film
The agent researched the real cold-brew market and the actual street address, then produced a logo, 16 images, four social ad variants, and a six-clip cinematic ad film with an original voiceover and music score.
15:35 – 16:40
14 · The honest label: it directs, other tools render
The published campaign page explicitly credits which external tools generated the video, images, and voice — the model's role was research, strategy, and direction, not the rendering itself.
16:40 – 16:59
15 · The guardrails came back up
Safety systems are now more aggressive and will silently hand off tasks to a lesser model when flagged as risky.
16:59 – 18:48
16 · The money: it's expensive as hell
Concrete pricing (~$10/M input, ~$50/M output tokens), free-usage windows for subscribers through July 7, and the real cost of running big jobs on pay-as-you-go credits.
18:48 – 22:21
17 · Where I land: the real deal, but a specialist
Verdict: genuinely capable of one-shotting full projects a prior generation could not, but slow, expensive, and best reserved for big, hard, high-value jobs rather than everyday tasks.
Atomic Insights
Lines worth screenshotting.
A single one-paragraph prompt with no follow-up instructions produced a working Next.js sales funnel, checkout flow, 11 generated images, 6 video ads, and cold outreach drafts in one unattended session.
The same one-shot prompt technique built a from-scratch Minecraft clone with its own chunk system, lighting engine, and day-night cycle in about ninety minutes.
A GTA-style open world with vehicle physics, traffic AI, pedestrians, and a working police wanted system was built from a single prompt and stress-tested itself with a 90-second five-star chase before handing it over.
When five self-review agents hit a session rate limit mid-review, the model didn't hide the failure — it reported the crash, then manually reviewed its own code and found four real bugs.
A team of five AI reviewer agents can all die simultaneously from the same session rate limit, since they share one usage pool rather than running independently.
The model debugged a UI complaint about an unplayable embedded game preview by realizing the preview window was stealing mouse capture, then told the user to open it in a real browser tab.
Premium reasoning-tier tokens cost roughly $10 per million input and $50 per million output — about double the mid-tier model's pricing.
A single large build on the highest reasoning setting can cost tens of dollars in pay-as-you-go credits once free weekly usage is exhausted.
One creative-director prompt run inside a multi-agent tool produced a full brand campaign — logo, 16 images, a cinematic ad film with an original voiceover and score, a pitch deck, and a published page — in about thirty minutes before hitting a session cap.
The model planned and gave an itemized cost estimate for a marketing campaign before spending anything on generation, instead of spending first and reporting after.
When a cinematic drone shot failed to generate, the model fell back to a real Google Street View capture of the actual business address and used that instead.
The published campaign page included its own honest disclaimer, labeling exactly which parts were generated by which external tool (video model, image model, synthesized voice).
The safety system on this model tier was tightened after a temporary shutdown, and it will silently hand off tasks to a different, lower-tier model when it judges a request risky — producing answers that can feel a notch weaker without explanation.
A four-hour, multi-build session across one day consumed several million tokens total, illustrating how fast heavy agentic use burns through a weekly usage allowance.
Running a planning model paired with an execution model, then having the planning model review the finished work, is presented as a cost-effective hybrid pattern for expensive one-shot jobs.
Takeaway
One long, expensive prompt can now replace a small team's day of work.
WHAT TO LEARN
The highest-reasoning tier of a coding-capable AI model can one-shot an entire finished deliverable from a single detailed prompt, but only pays off on hard, high-value jobs, not everyday small tasks.
A single, detailed one-paragraph prompt with no follow-up guidance can produce an entire finished deliverable — a working sales funnel, a game, or a marketing campaign — not just a draft or outline.
Long autonomous AI runs will hit usage limits mid-task; the value is in whether the system resumes exactly where it left off rather than restarting from scratch.
When an AI system reports its own failures honestly (a bug, a dead review team, a missing capability) instead of hiding them, that transparency matters more for trust than never failing at all.
Multi-tool AI orchestration (one model reasoning across research, image generation, video generation, and outreach tools) is now capable of assembling agency-grade marketing deliverables in under an hour.
Premium reasoning-tier usage is roughly double the cost of a mid-tier model, so treat it as a tool for a small number of high-value jobs per week rather than a default setting for routine tasks.
Clear labeling of what a model generated itself versus what other connected tools rendered is a trust signal worth demanding from any AI-produced deliverable, not just a nice-to-have.
A cost-efficient pattern is to use the most expensive model only to plan, orchestrate, and review, while a cheaper model handles the bulk of execution in between.
Safety guardrails on capable models can silently downgrade a task to a lesser model when the system judges it risky, which can explain an unexpectedly weak answer with no visible warning.
Glossary
Terms worth knowing.
UltraCode / highest reasoning setting
The most compute-intensive reasoning mode available for a coding-focused AI model, used for long, complex, unattended jobs at a much higher token cost than default settings.
Context compaction
A process where a long-running AI session compresses or discards earlier parts of its working memory to fit within its context limit, which can cause it to lose track of previously assigned task identifiers.
Rate limit / usage cap
A ceiling on how much of a paid AI service a user can consume in a given period (weekly or per-session); hitting it pauses or blocks further work until it resets or the user pays for more.
Diarization / speaker tagging
Not used in this video — omit if not referenced.
Multi-agent orchestration tool
A platform where a team of AI agents, each running on its own cloud machine with its own browser and tools, carries out real work autonomously rather than just responding in a chat window.
One-shot prompt
A single, complete instruction given up front with no follow-up guidance, requiring the AI to plan and complete an entire multi-step project from that one input alone.
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.
17px
metaphoranalogy
00:00Fable five is back, and I'm not going to sit here and read you benchmarks because, honestly, it is all just noise. You can watch 10 of their videos for that specifically, and I'd rather just show you what this thing actually builds. Here's what I did.
00:10The second that it actually came back online, I gave it four different prompts, one prompt each without any hand holding, no step by steps, and I ran every one of them on UltraCode, which is the highest setting, the one that eats your usage alive. And I blew past my limit about three times overdoing this. Now two of these are going to be real businesses, and two of them are just games that I built just to see if it would break.
00:30What I found is that it didn't, so let's get into it because the first one still kinda messes with my head. But real quick, just so you know what you're actually watching, Fable five, it's Anthropic's most capable model right now. It's built for the long, the messy, big jobs, the kind of thing that you used to break into 20 little steps.
00:44Now with this, you hand it one goal and you walk away. And beyond that, Fable five got pulled offline just a few weeks back where the US government just stepped in over security. So, Anthropic, they just had to take it down for everybody.
00:55So they tightened it up, and they came back online yesterday. And I'll get into what that actually means for your usage later. For now, it's back, and that's the whole reason we are here.
01:02That's the test. Can it take one paragraph and give me back a finished thing? Not just a demo, an actual finished project.
01:08But first up is the one that I would actually put in front of a business owner. It's just a complete go to market machine, and I do mean complete. So the product that we're using, it's completely made up, but the build, it's completely real.
01:17And what I did is I just invented a little SaaS called Ringback AI. Effectively, it's just an AI that calls back every missed call for local service businesses, even plumbers, HVAC, dentists, so they just stop losing jobs to voice mail. So that part is the pretend part.
01:31Here's the prompt that I actually gave it. It'll be a complete all round go to market machine for the product in the brief below from scratch in one project. Treat this like I handed you the product and said get it to market this week and start filling the pipeline.
01:42You are the strategist, the developer, designer, video producer, and the outbound growth agent. Build a funnel as real code, generate every image and video for real through the connected MCPs, and run real cold outreach. And lastly, said make decisions yourself and state your assumptions instead of asking me.
01:57By the way, I'll have all of these resources, the guide, the step by step, and plenty of other AI resources that we do not put on my YouTube channel available for completely free inside of my school community. Check that out. Link will be down below in the description.
02:07Beyond that, I then just gave it a one page brief and told it which tools to actually use. I also said to generate the images and the video ads for real through Arcadz, build the outbound list for real through Clay, write the cold emails as real Gmail drafts, and the last line of my prompt was this is on Ultracode.
02:23Go all the way. So it's it. Now from here, it just built the whole funnel as a real Next.
02:27Js site. So there's about six pages. We have a long form sales page.
02:31We then have this free report opt in. We have a checkout with Stripe in the test mode, an order bump, an upsell, even a thank you page. And this isn't just some wireframe junk with fake text.
02:42So if you actually read it, the headline, never lose another leads to a missed call. We have a real stats bar right here and under sixty second callback, 8,400 a month, all recovered, and there's even a fake product dashboard mocked up right in the page.
02:56We have Plumber Joe's 47 missed calls caught, 31 jobs booked, and it actually wrote some testimonials. Here it says first week, it booked a $1,400 compressor job at 08:45 on a Friday night.
03:07It even wrote its own honest disclaimer at the bottom that the 8,400 number is just an average, not a guarantee, and I didn't ask for that. It just has some taste to it. Now the lead magnet that it promised here, it didn't even fake that part.
03:18It generated a real eight page PDF guide in code, and then we have the images. So every picture on this site, it's completely generated. It made 11 of them through Arcadz.
03:27We have a logo. We have a transparent version of the logo. They hear a shot of the technician holding the phone, three application mock ups, and four different ad creatives, And it also QA'd them itself for the text that came out garbled.
03:39Next up, we see that it rendered a six video ads, and these are just actual UGC style ads with an AI actor just talking to a camera with a voice over, the captions burned in, and then the part that actually made me sit back and was kinda blown away is the outreach. So here, it just used Clay to find the real business owners in 30 of them, verified them across Orlando, San Antonio, Austin, Charlotte, and even Denver.
04:00And then it just wrote a personalized cold email to each one, and it dropped them all into Gmail as drafts. If So you take a look at these, every single one is completely different. But check this one out right here.
04:09It says, hi, John. Three decades of BERT roofing. Hi, Julian.
04:13Your NATE certified crew covers the Charlotte Metro. It's reading just who they are and writing to them. Let me just be a real judge here because, obviously, that's the whole point of this video.
04:21This is not perfect, and I wanna be completely straight about all of the rough edges that we see here. We're halfway through. It hit rate limits, and I had to just rekick the prompts twice.
04:29And at one point, it lost some of its own task IDs to context compaction, and I just had to redo a batch, which obviously is annoying. And the honest piece is those cold emails, it didn't blast strangers because I told it verified emails only, and my Clay email lookup wasn't actually switched on at that point. So it played it safe, and it addressed all the drafts to my own inbox with a do not send note in each one of them.
04:50So it's just a loaded outbound engine that's just one toggle setting away from being live, not a machine that already emailed the entire world. That's the correct call, but you should know it. Let me put some real numbers on this because I do think that matters.
05:01So this one, it went for the better part of about two hours, even counting the couple of times that it had stopped, and I just had to nudge it. And it also chewed through close to a million tokens of context to actually do the job, and that's the thing to understand about this model. That's not a chatbot answering question in three seconds and getting back to you instantly.
05:19That's a machine just grinding on one goal, holding the whole project in its head for about two hours. Now let's just take a step back and look at what that actually is. So we came back with a funnel.
05:27We have the real copy. We have this PDF. We have 11 images, six video ads, and a warm outbound list from one paragraph in one afternoon.
05:36And I need you to just actually sit with this for a second because it's kind of insane. Like, there's not one other model on this entire planet that could pull all of that off from just a single prompt. Not one of them.
05:44I've tried all of them. This is the first time in my life that I've watched a model just do the whole job, start to finish by itself. Opus 4.8 and ChatGPT's 5.5 was moderately close, but not to this level.
05:54This This is genuinely a first of its kind, and I do not say that lightly. And if you do do this kind of work for clients, you know that's not a small thing. And I'm not gonna throw a crazy number at you, but a package like this, it's genuinely a few thousand dollars of work with just one prompt.
06:07And, again, I'm going to be dropping the full go to market prompt in the brief for free in my school community, so check that out below. Moving on. A lot of people, the first thing that they do with a new model like this is just throw a game at it, and I did the exact same thing in my last video.
06:18I had Opus 4.8, which up until now was the best model out there to just build me a Minecraft. So let me just actually pull that one up so you can see it.
06:25So that is the Minecraft Opus 4.8 had actually built me. And, honestly, for what it is, it's a solid little voxel sandbox where you can place blocks.
06:33You can even break blocks. It runs. Nothing is wrong with this.
06:37But remember, this is the model that everyone said was the king just a couple of days ago. Keep that in your head for just a second. Now here's the exact same job.
06:44Hand it to Fable five. It's just one prompt. I basically told it to build Minecraft from scratch and actually make it feel like the real thing.
06:50So here's what I said. Build Minecraft from scratch that runs in the browser. One prompt.
06:54No shortcuts. Write your own real world generation, your own chunk system, and your own block textures and code.
07:00Let me place and break blocks, add a day night cycle in a proper hotbar, just and make it actually feel like the real Minecraft, keep it runnable the whole way. But we actually call this a voxel craft, and the moment it loaded, I knew this was instantly a difference here. So if we look at the top left, this isn't just a world.
07:14It's actually a whole engine readout. So we have your exact coordinates. We have the chunk that you're standing in, the direction that you are facing.
07:22We have the light level, the in game time. It's got a day night cycle running as well. We have a world seed about six two five two seven four.
07:31It wrote its own chunk system and its own lighting engine from complete scratch. The terrain, it has real hills. It has overhangs.
07:38The trees, they genuinely look like trees, and there's a full block hop bar down the bottom with the grass. We have the dirt, the stone, the planks, logs, glass, everything. Most importantly, it plays right.
07:48So I can genuinely break. I can place things. I can walk around, and the lighting actually updates.
07:54Now if we compare things and put this next to what Opus 4.8 gave me, the model we all thought was the best, and it's not even a fair fight with this. So that one was just a nice little demo. This one feels like an actual game.
08:05So it's the same idea, one model generation apart, and the gap is actually pretty insane. I will say partway through the build, I hit my usage limit, the whole thing just stomped on me as you would expect. And that's the honest reality of running this stuff on UltraCode.
08:18These long, complex jobs, they're big enough that you will bump into your limits, but the part that actually matters is that it didn't lose its place. Once my usage actually just freed back up, I picked right back up where I left off, and it finished the world for me. So it didn't make me just start all over from scratch.
08:34So all in, once you count that pause, this Minecraft was about an hour and a half of work and a few 100,000 tokens built from a completely blank folder. Beyond that, there was a moment where I thought that it initially had failed when it first handed the game to me inside the little chat preview window. I literally typed, I can't play the game, and it just figured out why on its own.
08:53So the embedded preview, it it was just stealing the mouse capture. It told me to just open it in a real browser tab. Boom.
08:59It worked, and the game was just fine the whole time. The window was the problem, and it debugged my complaint better than I could have ever done. Okay.
09:05So if Minecraft was the warm up, this one is where I stopped believing my own eyes, where I asked for a GTA style open world. This is all just in one prompt. Here's the important part of what I told it.
09:14Go to GTA style three d open world game that runs in the browser, write the vehicle physics, the character controller, and AI yourself. So generate all meshes and textures procedurally in code.
09:24Do not copy any rock star assets. Walkable and drivable Neon City, day and night cycle, drivable cars with arcade physics, traffic and pedestrians, a wanted system with the police chase. Of course, we have to throw that in there, and some basic combat.
09:37So just keep it runnable after every step. Now if you notice, I told it to build everything itself and not touch any real rock star stuff. So every building, every car, every texture just had to be its own code.
09:47And it wrote back, cool. I'll build you a Vice City flavored district on 3.js. Then it just went.
09:52Okay. So this is Niang Bei. Just take a look at this.
09:54It built a full city. It's not just a street. It's not a block.
09:57This is genuinely a full city. So it's 12 blocks by eight. There's around 600 buildings, and I actually went and counted because I did not believe it.
10:05So we have the beachfront hotels. We have the glass towers downtown. We have parks, parking lots, palm trees, all down the roads, and a real sand beach, and an ocean that is actually moving.
10:15And now just watch the sky. There's a full day night cycle completely running. So the sun literally sets while you were driving, and the whole city flips over into neon that I'd asked for, all from one prompt, but that is genuinely insane.
10:27I'm gonna be saying that a lot, but that is insane. And this part right here, it's not just a pretty backdrop. This is an actual game.
10:33So you get in a car. Look at how it drives. Like, there are three different cars, and they all feel different.
10:38You've got a hand brake. You can drift the car. It leans into the corners, and it throws sparks when you actually clip something.
10:44And if you look around a little bit, there's traffic. So we have real AI cars driving lanes, stopping at the lights. We have people on the sidewalk that scatter when you drive at them.
10:53Then I found the wanted system, and this is where I just completely lost it. So you start causing chaos. And if you look at the top right, the stars, of course, if you've ever played GTA before, they start climbing, and up to seven police cars, they spawn in, and they hunt you across the whole map.
11:07They can ram you. They can chase you on foot. If you go down, it's gonna be a slow mo in the, you know, classic wasted screen straight out of GTA, and it drops you back onto a beach about, you know, a $100 lighter.
11:18Now the kicker with this is it tested itself. So it ran its own ninety second five star police chase as a stress test, and it came back with a rock steady 60 frames a second and zero errors with it. Again, it stress tested its own game before it even handed it to me.
11:32Come on. That's that's crazy. Let me just come down off the ceiling for a second because none of this is free, and I'm not going to pretend that it is.
11:37A big build like this on UltraCode absolutely torches your usage, and I could watch my limit just draining while it had worked. In this game alone, it ran for over an hour. It chewed through the better parts of a million tokens, and it ate a scary chunk of what I pay for even every single month.
11:52I'll give you the real pricing and the actual numbers at the end, but just know nothing that you're watching here is going to be cheap, but that's to be expected. Now the honest part about this is it tried to run a team of five AI agents just to review its own code, and all five of them died. So they hit the session rate limit, just mid review.
12:07So instead of pretending everything was completely fine, it wrote that down. It told me exactly what had happened, and then it went back, and it did the review by hand. And what it found is about four real bugs.
12:17So we have a memory leak where old cars just never got cleaned up. We have traffic turning into oncoming lanes because of just a flip sign in the map, destroyed cop cars that just weren't raising your wanted level, and it fixed all four of them, and it just told me that it fixed them. And earlier in the build, I'd also told it that the character couldn't walk, and it just traced it to the exact file in line, and it patched it while I just watched the entire thing.
12:36And that's the thing that I keep coming back to. It's not that it never messes up. It's that it tells you when it does, and then it goes and it fixes it.
12:43And that is completely different relationship with a model. So, yeah, a playable open world game, its own physics, its own city, from just one paragraph And a couple of days ago, that sentence would have sounded like a complete lie. But with Fable five, it's just a Tuesday afternoon.
12:56When people say anything is buildable now, this is the kind of thing that they mean. Check this out. It's not just me pushing this.
13:01Like, somebody posted this the other day. They asked Fable for an extremely delightful calorie tracker, one four hundred word prompt, and it got back this ridiculously polished, fully animated application that people lost their minds over.
13:12It's all over x right now, so this isn't just a me thing. This is just what one prompt does now. Alright.
13:17Now I wanna show you the fourth build, but I have to introduce something because a lot of you will not have seen it. But there's this other tool that I've been obsessed with lately. It's called HyperAgent.
13:24It's built by the Airtable team. Now this isn't an ad. I've genuinely been spending a crazy amount of time in this thing, so I'm gonna do a full review of it on its own soon, so I won't go too deep today.
13:33So for right now, just understand one thing. It's an absolute powerhouse. But the short version is it's just a team of AI agents that each run on their own cloud machine with a real browser and real tools so they'd actually go do the work, not just chat back and forth with me.
13:46Now I used to run it paired with Opus, and it was great, and I built real stuff with it. But now that Fable is back, I switched it over, and it's on a completely different level now. So I just wanted to give Fable the hardest marketing job that I could actually think of and let it cook inside of HyperAgent, and this is the one that you could genuinely be selling right now.
14:01With this, I just made up a fake coffee brand called Kaya, just a premium cold brew launching a flagship cafe on South Congress in Austin, and I gave Fable one prompt as the creative director. Here it is. So my creative director and video producer from this one prompt produced a complete launch campaign for the brand review headlined by a cinematic ad film using only your built in tools, so things like web research, image generation, all that stuff.
14:22No outside tools. Plan first, and then show me an itemized cost estimate before you spend anything. And then I just gave it some information about the brand, Kaya Coffee.
14:30Now the first thing that it did before spending any amount of money was planning the whole campaign out and just showed me an itemized cost estimate. So from there, it asked for nothing else, and I just built it. Now in about thirty minutes before it hit its session caps, it made 31 assets.
14:43And I will be honest with you, watching them render onto the canvas one after another after another was one of the most satisfying things that I've watched an AI do. Let me actually walk you through it because this is an entire campaign. Really quick, just think back to that go to market build that we had earlier.
14:56To pull that one off inside of Claude, I had to connect three separate tools by myself. So we had Arcadz. This is just for the images and the videos.
15:03We had Clay for the leads and the sourcing, Gmail for the outreach. And I actually just wired all three of them up by hand. In here, I connected nothing.
15:12So the research, the image generation, the video, the audio, the maps, the slides, all the documents, it's already built into the one agent. So Fable, it just reaches for whatever it needs at that time.
15:23So we have one brain, every tool already sitting in its hands. And that's an absurd amount of power in one place, and that's exactly why I cannot put this thing down unless I'm hitting my usage limits. So the work itself, it just did live research first where it pulled real numbers on the cold brew market.
15:35It's a $3,000,000,000 category growing about 17% a year. It looked up the actual neighborhoods of the South Congress rents, the foot traffic. It even looked up what the word Kaya means.
15:45So the strategy underneath all of this is real, and it's not just vibes. And then we have the images. So it made 16 of them, and these are not just stock looking throwaways.
15:53We have a logo, a word mark, a full color palette. We have photo real shots of the cup and the can on location plus a clean flat lay. So the branding that it landed on, Kaya, Rise Lo, on this warm, dark background.
16:04Honestly, this looks like a real coffee brand that you'd stop and look at at the street. And then four separate social ads, the square, the story, landscape, each with a different hook, ready to run tomorrow. And then we have the video, and this is where I really leaned into it.
16:16Six clips, and the centerpiece is just a thirty second cinematic ad film, and it didn't just spit out one shot. It generated the film in pieces, the slow pour.
16:24We have the steam coming off the top, the morning light through the window, the hero shot of the product, and a lifestyle beat with people in it. And then it just cut all of those together into one shot so you can actually see the individual clips sitting right there on the canvas. We have the pour, the feeling.
16:39We have all of it. Now watching a model like this direct its own little film shot by shot is a genuinely strange thing to sit with, though. The part that actually made me sit up was the location.
16:48So I told it the cafe is at 1011 South Congress, so it went and pulled up a real Google Street view of that exact address, navigated to the real block, and it worked into opening of the film. So it grounded a brand that doesn't even exist yet in a real place on a real street. I did not expect any of that.
17:03And if we turn the sound up for this, it wrote a voice over script. It generated the read. Kaya, Rise Lo, a full thirty second of it, and then laid an original music bed just all underneath that.
17:14So there's no stock track. It scored the thing itself. In the whole spot, it sounds like something that an agency is going to hand you after three weeks and, you know, a pretty sizable budget.
17:22On top of the film, it built a full pitch deck to just present the whole campaign and this one page creative brief, and then it published all of it as just one clean showcase web page, strategy, and its sources just sitting right there on the page, which brings me to the honest piece, and I actually love that I did this.
17:37Read the bottom of this page. It labels it itself. So it says straight up, the film footage was generated with Google's VO.
17:44The images with Google's Gemini, the voice over is synthesized speech. So let just be really clear about what actually happened here because this does matter. Fable five is not the thing painting the pixels.
17:55Fable's just the director here. It did the research. It wrote the strategy, wrote the script, the art directed every shot.
18:01It called the right tool for each job, and it just assembled the whole thing into a finished campaign with an honest label on it. So the intelligence, it's Fable, and the rendering is those other tools that I just mentioned.
18:12That's the honest picture with this. And, of course, it wasn't flawless. So that cinematic drone flyover over the address, that one actually failed on it.
18:18So the real street view shot that you saw earlier was its plan b, and it just ran into HyperAgent's thirty minute session cap, so it stopped itself before it was just completely finished. But look at what it made in that half hour. We have a launch film, 16 images.
18:31We have the videos, a pitch deck, the creative brief, and a live web page with this. And that's a package real agency bills a few thousand dollars for. So if you run a business or if you do this kind of work for your clients, that's the sentence that I would be sitting with for a second, just thinking how you can actually use this for fulfillment or how you can start a business offering this.
18:48I mean, this is insane. Alright. So you've seen what it does, but now I just wanna cover what you actually need to know before you go and use it because it did not come back exactly the way that it left.
18:56So it's live, again, everywhere, Claude, Claude code, Cowork, all of it. But there are two things that you have to understand going into this. First, the guardrails, they just came back turned way up, and this is the honest catch with this.
19:07So the new safety system, it's more aggressive. In Anthropic, they say themselves that they will flag normal coding and debugging more often than it should right now. So when it gets nervous, it's quietly going to hand your task to Opus instead.
19:19So if an answer ever feels a notch dumber than you expected, that's probably what happened. You didn't do anything wrong. And second, the money, because this thing is genuinely expensive as hell, and I want to say that plainly.
19:32So it's about $10 per million tokens going in, 50 going out, and that's double what Opus costs. And remember, all those numbers that I keep dropping throughout the video, close to a million tokens on, you know, the go to market build, better part of a million on the GTA one, and you add all four up. And I ran something like four hours and several million tokens throughout this model in a single day.
19:52But here's how the plan works right now. If you are on pro, if you're on max or team, you get Fable for up to half of your week usage for free, but that's only going to be through July 7. After that, it flips to just pay as you go credits.
20:03And once my free half was gone, those builds were costing me real money. And a single big one on UltraCode can run you tens of dollars a pop. And that is why I just blew through my limit three times making this video.
20:13So do not point this at your little task. You will burn your whole week in just one afternoon. So save it for the hard, the expensive one shot tasks where, you know, single good run pays for itself.
20:22That's the whole move. Even using a hybrid where you have Fable planning and doing all the orchestrating, then Opus actually doing the execution, and then having Fable just to review everything again.
20:30But where do I actually land after pushing this incredibly hard on just one day? I will say that this is the real deal. None of this is hype.
20:36The stuff that I built, a full go to market machine, a Minecraft engine, an open world game, even an agency grade launch campaign, like, are all things the last generation genuinely could not one shot. It's the first model where the bottleneck just stopped being the model, and it started being me. Like, how can I clearly describe what I want?
20:53But it's not a specialist. It's not your everyday driver. It's very expensive.
20:57It's slow because it thinks a lot, and the guardrails are very jumpy right now, I will say. So you don't hand it your quick little job. You have to hand it the big, scary, long ones, the complex ones that other models couldn't figure out, the whole application, the whole campaign, migration that you've been dreading.
21:11And on those, it pays for itself in a single good run. Now if you run a business, that's the real headline. This isn't a toy.
21:17Two of the four things that I built today, they're just things that you could put a real price tag on, and the gap between people who point this at real work and people who do not is about to get very wide. If I were you, I would get comfortable with it now while half of your free usage is still on the house. So with that being said, that's everything that I wanted to be covering.
21:32Go check this out for yourself while you still can, and this is genuinely the best thing that I have ever seen. Point it to your hardest jobs. Figure out whatever you've been trying to figure out to get to that next scale, that next goalpost, whatever it may be.
21:46Now if you are a business owner looking to implement solutions just like this or anything else to scale your business in 2026, then you can book on call with our team at Reprise. We'll go through a free audit and show you where you can leverage AI and how we can help you increase your bottom line this year.
22:00And if you're looking for all of our other free AI resources, then make sure to join our free school community. Link will be down below in the description. But with that being said, drop a comment how you guys are using this, what your thoughts are on everything, hold a bottle of them just lifting the band, if it's been switching up on you and dropping to Opus.
22:14Genuinely curious because I haven't found that to happen too much myself. With that being said, thank you guys for watching, and I'll see you in the next video.
The Hook
The bait, then the rug-pull.
Instead of benchmarks, the creator hands a newly-reinstated top-tier AI model four one-paragraph prompts and walks away — then comes back to a real sales funnel, a working Minecraft clone, an open-world driving game, and a full ad campaign, each built without any further hand-holding.
Frameworks
Named ideas worth stealing.
20:34concept
Hybrid planning/execution/review pattern
Top-tier model plans and orchestrates
Mid-tier model executes
Top-tier model reviews the finished work
A cost-saving pattern for using an expensive high-reasoning model only for planning and final review, with a cheaper model doing the bulk execution in between.
Steal forany workflow where an expensive premium model would otherwise be run end-to-end on a long task
CTA Breakdown
How they asked for the click.
VERBAL ASK
22:12link
“Check that out below... link will be down below in the description. Book a call with our team at Reprise for a free audit.”
Soft repeated plug for the free Skool community throughout, with a harder pitch for a paid AI-consulting audit call near the very end.
A 14-minute walkthrough of the Noose desktop installer that finally lets non-technical users run one of the most capable open-source AI agents without touching a terminal.
00:24
00:35
00:44
00:56
01:04
01:21
01:59
02:15
02:35
03:08
03:27
03:39
03:52
04:20
04:36
04:57
05:48
05:57
06:06
06:25
06:42
06:56
07:11
07:33
07:54
08:29
08:43
09:05
09:30
09:52
10:03
10:20
10:37
10:54
11:27
11:44
11:58
12:18
12:36
13:16
13:32
14:45
15:21
15:41
16:26
16:46
17:07
17:27
17:46
17:57
18:12
18:34
18:49
19:11
19:16
19:52
20:05
20:31
20:36
21:03
21:11
21:54
22:05
24:57
14:27
16:24
27:40
16:58
14:31