What Can You Build with Claude Fable 5.2? Hardware, Games, Costs
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I go through the best things people built in the first week of Claude Fable 5.2: circuit boards, robots, browser games and short films. I also cover who would pay for this kind of work, what it costs next to Opus 5.5, and three recipes to try one yourself. It's for anyone building with AI who wants to know if Fable is worth the price. Every cost and claim from a builder is theirs, and I say whose.
What I cover in this video
- 0:00 Hardware, games and money from one model
- 0:47 Pricing and what changed in Fable 5.2
- 3:06 Hardware: three devices from one paragraph
- 5:24 Software, games and film
- 7:41 Money: four ways to get paid, and the catch with each
- 10:44 Is it worth it? Fable vs Opus costs
- 13:01 Three recipes to try one yourself
- 14:33 Why the blueprint walkthrough is the one I'd copy
What changed in Fable 5.2
Anthropic released Fable 5.2 on November 9 on all paid plans. It costs $8 in and $40 out per million tokens, down from $10 and $50 for Fable 5.1. Opus 5.5 is $4 in and $20 out, so per token Fable is exactly double.
Three things explain why people started building physical stuff with it:
- Long runs. Anthropic says one engineer had Fable build a prototype operating system from an empty folder, using over a thousand sub agents over 8 days, and it boots on a laptop.
- 3D and CAD. On BenchCAD it scores 83.5 versus 65.3 for Opus 5.5 and 44.4 for Fable 5.1. On PCB Gym, a circuit board test, it went from 25% with Fable 5 to 69%.
- Robots. On RoboDojo, a simulated two arm robot test, it went from 27.6% with Fable 5.1 to 64.9%.
Those are Anthropic's own numbers. On the independent side, Artificial Analysis has Fable 5.2 at number one out of 80 models on its Intelligence Index, 65 versus 59 for Opus 5.5. On its Coding Agent Index the two are tied at 58.
The builds: hardware, games and film
Hardware is real, but early
Anthropic and a hardware company called Diode gave Fable a one paragraph prompt and got back three devices: a voice assistant hub, a USB speakerphone and a handheld virtual pet. Fable wrote the spec, picked the parts, routed four layer boards, designed the cases and programmed the firmware. Anthropic says each board was ready for manufacture in under a day and all three worked without electrical modifications. The caveats: Anthropic's own write-up calls these early prototypes, Diode's engineers did the soldering, and Anthropic has a business relationship with Diode.
For an outside test, the builder behind Zack's Lab gave it the spec for a high current power board that took weeks to design by hand. It got stuck with 53 connections it couldn't route and 39 spacing errors after $417 of credits. After running all day, the builder says it fixed all of that, but the hand-routed board still packs in more power. So a small gadget board, yes. A serious power board, not quite yet.
On the robot side, one builder designed a full humanoid overnight with CAD written in code using build123d and tested in the MuJoCo physics simulator. Nothing was built. Another ran a real robot arm through a 62 move game of checkers, where it missed 14 grabs along the way.
Software, games and film
The build everybody copied came from Felix Rieseberg, an engineering lead at Anthropic. One prompt with no human feedback produced a full first-person shooter with no images or 3D models. Every texture and model is code. It ran for eight hours and Felix puts it at about $550 at API pricing, with about 2.7 million tokens out and 445 million in. The input number is so much bigger because an agent rereads its own project over and over. Felix says the run was heavily cached, and cached input costs 20 cents per million instead of $8.
In film, Riley Brown gave it one prompt live on stream. It wrote the code that builds a room in Blender, rented 10 GPUs on RunPod and rendered it in 4K in 52 minutes. Riley says the render cost $5.75.
Four ways to get paid, and the catch with each
The Remote Labor Index is 240 real paid freelance projects worth over $140,000 of work. In October 2025 the best AI could do 2.5% of it. Anthropic says Fable 5.2 does 41.3%, versus 21.3% for Opus 5.5.
- Blueprints into 3D walkthroughs. Everest Chris says these go for $5,000 to $10,000 a plan. I checked the market: studios charge from about $800 for a basic 360 tour of an apartment, and walkthrough animation starts around $2,200 per finished minute. The catch is measurements, so check every room by hand.
- Interactive learning sites. Someone built a free 3D electronics course with eight real boards. I couldn't find anybody showing a paying customer yet, so it's an opportunity and not a proven price.
- Hardware prototypes for inventors. Have a real engineer check the board before anybody pays for manufacturing.
- Game prototypes for indie studios. Eric built a GTA style open world prototype in 2 hours and 14 minutes on a $200 plan and says a studio would quote $90,000 and 8 months. That's Eric's number. One studio's price guide puts a playable slice at about $8,000 to $30,000. It's a prototype a studio could pitch with, and you can't ship it as a game.
You'll also see posts about a high schooler making $4 million a month with Fable. There's zero proof behind those.
Is Fable worth it next to Opus 5.5?
I found eight Fable versus Opus side-by-sides. Fable cost more in every one, from about 1.4 times to more than 6 times as much, and it was faster in only two. In half of those posts, the people running them still said Fable's result looked better, so you're mostly paying for quality.
Artificial Analysis ran Fable inside Claude Code. On High it scored 58 at $7.51 a task. On Xhigh it scored the same 58 for $15.47. On Max it cost $44.22 and scored lower, 56.
The cheap trick is Anthropic's advisor mode, where Opus or Sonnet does the work and asks Fable on the hard parts. Charlie Hills had it design a robot that solves a Rubik's cube: Opus alone cost $2.88, Opus with Fable as the advisor cost $3.88, and Fable alone cost $7.25.
My rule of thumb: Opus 5.5 stays my daily driver. I pull out Fable for long jobs, anything 3D, CAD or spatial, and the problems Opus can't crack. I run it on High, and I use the advisor setup whenever I can.
Three recipes to try one yourself
These come from what the builders said they used.
- A browser game or world. Open Claude Code with Fable 5.2 and ask for a single HTML file using three.js with everything made in code, no images and no 3D models.
- A 3D scene or short film. Have Claude Code write Blender Python so the code builds the scene. For the Riley Brown route, give it a way to rent GPUs on RunPod and let it render.
- Hardware. Describe what the device does in one paragraph and have it lay out the board in KiCad and the case in code CAD like build123d. Have someone who knows electronics look it over before you send files to a board maker like JLCPCB.
The biggest builds ran for hours. Felix's ran for 8 and Riley's film took about 4 and a half. Set a spending limit on your API key before you walk away.
Key takeaways
- Fable 5.2 jumped on long jobs, 3D and CAD, and the physical world. On coding agent work it's tied with Opus 5.5 on Artificial Analysis.
- Hardware builds are real but early. Get an engineer to check any board before paying for manufacturing.
- Fable cost more than Opus in all eight side-by-sides I found. Run it on High, skip Max, and use advisor mode to cut the bill.
- If I had to copy one build, it's the blueprint walkthrough. Those buyers already pay studios for 3D tours today, and you can make the sample before you ask anyone for money.
FAQ
How much does Claude Fable 5.2 cost?
Fable 5.2 costs $8 in and $40 out per million tokens, down from $10 and $50 for Fable 5.1. Opus 5.5 is $4 in and $20 out, so per token Fable is exactly double.
Is Claude Fable 5.2 worth it over Opus 5.5?
In the eight side-by-side tests I found, Fable cost more every time, from about 1.4 times to over 6 times as much, and was faster in only two. I keep Opus 5.5 as my daily driver and use Fable for long jobs, 3D, CAD or spatial work, and problems Opus can't crack.
Can Claude Fable 5.2 design a circuit board?
Anthropic says Fable 5.2 designed three working device prototypes with Diode from a one paragraph prompt, though its own write-up calls them early prototypes. In an outside test from Zack's Lab on a high current power board, the hand-routed board still packed in more power, so have someone who knows electronics check the design.
What is the cheapest way to use Claude Fable 5.2?
Run it on High effort and skip Max, and use Anthropic's advisor mode where Opus or Sonnet does the work and asks Fable on the hard parts. In Charlie Hills' test, Opus with Fable as the advisor cost $3.88 versus $7.25 for Fable alone.
Chapters
- 0:00Hardware, games and money from one model
- 0:35What changed in Fable 5.2
- 2:26Night Crew Studio (my app, quick break)
- 2:50Hardware: three devices from one paragraph
- 5:31Software and games
- 7:39Money: four ways to get paid
- 10:43Is it worth it? Fable vs Opus costs
- 13:05How to try one yourself
- 14:21The one I'd copy
Transcript
Read the full transcript
0:00Anthropic says Claude Fable 5.2 just designed three working hardware prototypes from one paragraph. One prompt got an Anthropic engineer a shooter game, every model and sound from scratch. And one builder says 3D walkthroughs like his go for 5 to 10,000 a plan. But demos don't pay bills. So what do these cost and who's paying? So in this video I'm covering what changed? hardware, games, money, if it's worth it, and how to try one yourself. I went through hundreds of these builds so you don't have to. Let's jump right into it. So, Anthropic released Claude Fable 5.2 on November 9th, and it's available on all paid plans. It costs $8 in and $40 out per million tokens, down from 10 and
0:4750 for Fable 5.1. Now, Opus 5.5 is $4 in and 20 out. So per token, Fable is exactly double Opus. Keep that number in your back pocket. It comes back later. So why did people suddenly start building physical stuff with it? Three things explain it. First, it can run for a really long time on its own. Anthropic says one engineer had Fable build a prototype operating system from an empty folder using over a thousand sub aents over 8 days, and it boots on a laptop. Second, it got way better at 3D and CAD. On a CAD benchmark called Bench CAD, Fable 5.2 scores 83.5 versus 65.3 for Opus 5.5 and 44.4 for Fable 5.1. That's nearly double in one
1:34version. And on a circuit board test called PCB Gym, it went from 25% with Fable 5 to 69%. And third, robots. On Robo-Dogjo, a simulated two arm robot test, Fable 5.1 scored 27.6%. 6%. Fable 5.2 scores 64.9. So more than double, which is insane. And those are Anthropic's own numbers, but the independent ones are in, too. As of this week, Artificial Analysis has Fable 5.2 at number one out of 80 models on their intelligence index, 65 versus 59 for Opus 5.5. But on their coding agent index, the two are tied at 58. So take the hype for what it is. So on paper, it jumped on exactly the hard
2:19stuff. Long jobs, 3D, and the physical world. And the first thing people did with that was make hardware. Blueprint, Clay Studio, Chalk Talk, Neon Noir, Retro Arcade. Every one of these came out of the same app. Met Crew Studio has 14 looks built in. And if none of them fit your brand, you describe your own in plain words, and it builds it. So, you pick a look, you give it a topic, and it makes the whole short on your Mac on your own Claude plan. Nightcrew studio. Join at nightcrew.studio. So, let's start with the post that got over 7 million views. Anthropic and a hardware company called Diode gave Fable 5.2 a one paragraph prompt, and it came back with three devices: a voice assistant hub, a USB speakerphone, and a
3:06handheld virtual pet. Each one has its own board, a 3D printed case, and its own software. So, how does a model that has never held a soldering iron pull that off? Fable wrote its own spec for each device, and then it picked the parts, and then it routed four layer circuit boards, which is basically drawing every copper path between the parts, so none of them touch. For the hub's long paths, Claude even wrote its own router in about half an hour. And then it designed the cases and programmed the firmware. Anthropic says each board was ready for manufacture in under a day with minimal human input and all three worked without needing any electrical modifications. Now, before you get too excited, Anthropic's own writeup calls these early prototypes. Diodes engineers did the soldering,
3:53Claude did the design, and Anthropic has a business relationship with diode. So, take that for what you will. And here's a real world test from someone outside Anthropic. The guy behind Zach's Lab gave Fable and diode the spec for a high current power board he'd already designed himself over weeks. He says the design choices were remarkably similar to his. At first, it got stuck with 53 connections it couldn't route and 39 spacing errors after $417 of credits. He left it running all day, and he says it fixed all of that. But his own handout board still packs in more power. So, a small gadget board, yes. A serious power board? Not quite yet. Then the robots. Jake designed a full humanoid robot overnight. It wrote the CAD files in
4:39code with build 123D, tested it in a physics simulator called Mujo, and then made a sizzle reel in Blender. Now, nothing was built. It's a design and a simulation, but in one night. A CAD company called Atom had it lay out a Tamagotchi style circuit board in Keycad, build the case in Inventor, and render it in Blender. So, for actual hardware moving in the real world, Karolina had Fable run a real robot arm through a 62 move game of checkers. She says it noticed the arm sagging about 7 mm under load and corrected for it. It also missed 14 grabs, but a lot of the time she just said, "Try again." And it fixed it. And in Anthropic's own demo, Robot Arms solved a puzzle cube, unscrewed a bottle cap, and drew a
5:24picture of a robot. So, hardware is real, but it's early. software is where it went completely nuts. So, the one everybody copied came from Felix Reeseberg, an engineering lead at Anthropic. He gave Fable 5.21 prompt with no human feedback and got back a full firsterson shooter, and there are no images or 3D models in it at all. Every texture and every model is code. Claude didn't draw anything. It wrote shaders. It ran for eight hours and he says it came to about $550 at API pricing with about $2.7 million tokens out and $445 million tokens in. And those two numbers explain the cost. So why is the input number so much bigger? An agent rereads its own project over
6:10and over while it works. So the tokens going in pile up way faster than the tokens coming out. Felix says his run was heavily cashed and cashed input costs 20 cents per million instead of $8. That's why it landed in the hundreds of dollars and not the thousands. That post has over 4 million views and it turned into a challenge. Grums built his own version in about three hours in two stages using his own engine. So, there's way more. Blendy made a game that looks hand painted in under an hour, and he says every frame is about 1 and a half million brush strokes running at 200 frames a second in a browser tab. Kevin no made a boat ride through eight dreamlike worlds you can play in your browser. Then there's film where the views went crazy. And Schwan made a 3D
6:55film explaining Zeno's paradox, why your fingers never really touch, and it's at over 24 million views. That's the most viewed build of the whole week. Riley Brown gave it one prompt live on stream, and it wrote the code that builds an entire room in Blender. And then it rented 10 GPUs on RunPod and rendered the whole thing in 4K in 52 minutes. He says the render cost $5.75. A few days later, his two-minute film, The Fable of Claude, had it start seven copies of itself and rent 89 GPUs for a little over $20 in render costs. So, at this point, you're probably thinking, "Cool demo. Next." And that's exactly where everybody else's video stops. So, who would actually pay for any of this? So, first, the one number that made me
7:41take the money question seriously. There's a benchmark called the Remote Labor Index. It's 240 real paid freelance projects like game development, design, and architecture worth over $140,000 of work. Back in October 2025, the best AI could do 2.5% of it. Anthropic says Fable 5.2 does 41.3% versus 21.3 for Opus 5.5. So, the jobs people pay freelancers for are exactly where this jumped. So, here are four ways to make money with these builds. Who pays? And the catch. Number one is blueprints into 3D walkthroughs. Chris, who posts as Everest Chris on X, shared a prompt that takes a flat 2D floor plan
8:27and turns it into a 3D home you can actually walk through. He says these go for $5 to $10,000 a plan. So, I checked the real market. Studios charge from about $800 for a basic 360 tour of an apartment, and walkthrough animation starts around $2,200 per finished minute. So, people do pay for this. And here's who. Home builders selling homes before they're built. Architects and remodelers. And here's how I'd start. Pick one niche in one city, like kitchen remodelers. Make a free sample from one of their own floor plans and send it to them before you ever ask for money. The catch? A builder needs the measurements to be right. So, check every room by hand before it goes out. Number two is interactive learning sites. Someone
9:12built a free 3D electronics course with Fable. Eight real boards like an Arduino and a Raspberry Pi where you place the chips, solder them, and blink an LED. And he says, "Every 3D model in it is code." Who'd pay? Companies that sell kits, schools, and anybody with a product that's hard to explain. The catch, I couldn't find anybody showing a paying customer for this yet. So, that's an opportunity, not a proven price. Number three is hardware prototypes for inventors. An appliance manufacturer in India posted that work that used to take weeks of back and forth with suppliers and design houses can now happen in hours. That's just his take. So if you know a little electronics, you could design the first prototype board for someone with an idea. The risk is Zach's
9:57power board. Have a real engineer check it before anybody pays for manufacturing. And number four is game prototypes for indie studios. Eric built a GTA style openw world prototype in 2 hours and 14 minutes on a $200 plan. He says an indie studio would quote $90,000 and 8 months for that. That's his number. One studio's own price guide puts a playable slice of a game at about $8 to $30,000. The limit, it's a prototype a studio could pitch with, not a game you can ship. You'll also see posts like a high schooler making $4 million a month with Fable. There's zero proof behind those. Come on now. Don't take business advice from a screenshot. So, there's money here, but there's one problem, and it's the price. So,
10:44remember, per token, Fable is double Opus, but in real builds, the gap can get way bigger. Envil Dev built the same Cathedral of Ice with both. Fable cost $81.10 and Opus cost $45.97. Now, Fable did finish about an hour and 15 minutes faster. Then Mr. Tactical X built a glass pool. Fable cost $57.87 and took 90 minutes. Opus cost $14.39 and took 68. Now this is the part I didn't expect. I found eight of these Fable versus Opus tests and Fable cost more in every single one from about 1.4 times up to more than six times as much. And it was faster in only two of them.
11:29But in half of those posts, the people running them still said Fable's result looked better. So you're mostly paying for quality. And the independent numbers show the same cost problem. Artificial analysis ran Fable inside Claude code. On high effort, it scored 58 at about $75 a task. On extra high, it scored the exact same 58 for over $15. And on max, it cost over $44 a task and scored lower, 56. So cranking the effort all the way up just burns money. So here's the cheap trick. You don't have to run Fable for the whole job. Anthropic has an advisor mode where Opus or Sonnet does the work and asks Fable on the hard parts. Charlie Hills had it design a robot that solves a Rubik's cube. Opus
12:16alone cost $2.88. Opus with Fable as the adviser cost $3.88 and Fable alone costs $7.25. And a company called Atomic Agent showed Fable as the boss with a local Quen model doing the building. They say that was $2.98 in 11 minutes versus $15.77 in 21 minutes for Fable alone. So, my rule of thumb, Opus 5.5 stays my daily driver, and that's coming from the guy who calls it Slopus. Even Felix, who made that shooter game, says he still often prefers Opus 5.5 for everyday work. I pull out Fable for the long jobs, anything 3D, CAD, or spatial, and the problems Opus can't crack. I run it
13:01on high, not max, and I use the advisor setup whenever I can. So, here are three recipes from what these builders said they used. Recipe one, a browser game or world. Open Cloud Code with Fable 5.2 and ask for a single HTML file using 3.js with everything made in code, no images, and no 3D models. That's the same shape as Felix's shooter and Blendy's painted game. Recipe two, a 3D scene or a short film. Have Claude Code write Blender Python so the code builds the scene. And if you want to go the Riley Brown route, give it away to rent GPUs on RunPod and let it render. And recipe three, hardware. Describe what the device does in one paragraph and have it lay out the board in Kyad and
13:47the case in CodeCad like build 123D. One guy designed a guitar pedal in Kyad like this and he said he's zero in hardware. His files were ready to send to a board maker like JLCPCB. Just have someone who knows electronics look it over first. So for all three two things, the biggest builds ran for hours, not minutes. Felix let his run for 8 hours and Riley's film took about 4 and 1/2. So set a spending limit on your API key before you walk away. And please don't be the guy using Fable 5.2 two to center a div. So, if I had to copy one build from this whole week, it's the blueprint walkthrough. And it's not because of the 5 to 10,000 he says these go for. It's because those buyers already pay studios for 3D tours
14:33today. And you can make the sample before you ever ask anyone for money. In my resale business, nobody buys anything they haven't seen first. And a home builder is the same way. So, that's a whole week of X builds in one video. A like is how YouTube knows to show it to more builders. So, if it saved you the scrolling, give it one. Which one would you put your time into? The hardware, a game, or the walkthrough business? Put it in the comments. Have a blessed day.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Claude Code Examples: 20 Real Projects People Are Building
I count down 20 real things people built with Claude and Claude Code: games, hardware, apps, and agents that run a business. Number one is Dave Jones of EEVblog, who says Claude Code rebuilt his 121GW multimeter firmware in plain C from a single prompt in under 24 hours.
Claude Opus 5 vs Fable 5: Cost, Speed, and 3 Real Build Tests
I ran Claude Opus 5 and Fable 5 on high effort through three builds: a Three.js MMO, a Remotion video, and a landing page rebuild. Opus 5 won the game and the UI test, the video was close to a tie, and Opus cost about half as much. It's my new daily driver, but Fable 5 is much faster.
Vibe Coding Apps That Make Money: How I Reached $800 MRR
After about 500 hours of vibe coding, I launched Easy Flip, an AI eBay listing app that uses GPT Vision and the eBay API, and grew it to $800 in monthly recurring revenue. The approach was to start with an audience, plan distribution early, and fix bugs fast with Claude Code and Vercel logs.
