Yesterday a pile of demos dropped that I keep replaying in my head. Step 2 of the workflow below is the twist, and it’s the opposite of what most of us do with prompts. The creator, Matthew Berman, had early access to Anthropic’s new Fable 5.2 and threw nearly every project he could think of at it.
He says it’s the best model he’s ever used, “by far.” After going through his demos, I see why. He also points out where it falls short, which is just as useful.
What’s new
Fable 5.2 is Anthropic’s newest top model. The author ran it on games, desktop apps, physics sims, 3D renders, motion graphics and even a real 3D print. Here are the standouts:
- A Gears of War E-Day clone in Unreal Engine. It’s one playable level with cover, AI squadmates, revives and the timed reload. It used about 8 billion tokens, which he puts at roughly $11,700 to $15,000 at API prices. It took a few days.
- A full Excel clone in the browser. It has working formulas, a sample sales tracker and a real pivot table. The build took almost 50 hours, and he says he couldn’t find a single missing feature.
- A DaVinci Resolve-style video editor called Kestrel. It runs locally on Windows with waveforms, clip splitting, magnetic snapping, keyframes and keyboard shortcuts. Someone on his team already edits real videos with his own version.
- A Fall Guys parody called Tumble Beans. It has 10 levels, different game modes and working multiplayer. That run used over 9 billion tokens and about 71 hours, roughly $13,000 in API terms. His team played it live, and it got loud.
- Tokyo in Unreal Engine. It has traffic, crowds, trains pulling into stations and ray-traced reflections. That cost about 6.72 billion tokens and 70 hours, or around $11,000.
The twist
Here’s what surprised me. The prompts are tiny. The Gears of War prompt was three lines: build the most accurate clone you can, one level only, and make it look and feel like the real game. The Fall Guys prompt said “build 10 levels” and make them fun and varied. He only added multiplayer later.
The author says he likes to “under-specify” his prompts on purpose. He found the model very autonomous. It made calls for him, and he says it got them right about “nine times out of 10.”
The second surprise is the price spread. The giant builds cost thousands of dollars in tokens. Plenty of the smaller demos were cheap, though:
- Bullet through a water balloon. It’s a ray-traced physics sim in one HTML file using WebGL2, with no libraries. It has sliders for bullet speed, caliber, gravity and sun direction, and you can even see purple light through the balloon’s shadow. Cost: 41 million tokens, about 2.5 hours, $91.
- A 3D-printable tower under 10 grams. He gave it his printer (Bambu Lab P2S), the 0.4 mm nozzle and the PLA filament as constraints. The print held two MacBook Pros, an iPad Pro, an Asus laptop, a Mac Mini, an Nvidia DGX Spark, a reMarkable tablet and finally a power drill. Cost: about $37.
- Motion graphics explainers. A 1:52 video on software decompiling cost about $25 and took 45 minutes. A second one walked through how LLMs work: embeddings, transformer layers, training loss and next-token prediction.
- A dream house in Blender. It came with textures, a pool, reflections and glass, all for about $30. The only mistake he spotted was a tree in the middle of the driveway.
Not everything worked. A recreation of the Tesla Roadster event based on photos was, in his words, “the first bad one.” I liked that he showed it anyway.
Step-by-step mini-workflow
Here’s the pattern I pulled from the expert’s tests. You can use it today:
- 🎯 Pick a known reference. Ask for a clone or parody of something that already exists, like Excel, Fall Guys or DaVinci. The model already knows what “complete” looks like.
- ✂️ Keep the prompt short and set the scope. Say what you want and where it stops, like “one level” or “10 levels.” Let the model handle the details.
- 🧱 Add hard limits when they matter. For physical or technical work, list the real limits: weight, printer model, “no Three.js,” “single HTML file.” This is where the tower and balloon demos shined.
- 🔁 Extend in a second pass. Get a working base first, then add one big feature, like multiplayer.
Pro tips
- Watch your token budget. “Ultra Code” mode spins up many sub-agents at the same time. The author says it’s very expensive and doesn’t recommend it for everyday use. Save it for big jobs.
- Run a full code review. He pointed max thinking effort plus Ultra Code at his codebases and asked for a full review. It found bugs, security issues and small UI glitches. Jobs he expected to take 60 to 120 minutes finished in about 10.
- Know the pricing. Fable 5.2 is about 20% cheaper than 5.1: $8 per million input tokens and $40 per million output. Fast mode gives about 2.5x speed at $16 and $80.
- Don’t drop Opus 5.5 yet. The author says Fable 5.2 is clearly better, but Opus 5.5 is faster and much cheaper. Use Fable for the hard stuff and Opus for daily work.
The benchmarks back up his hands-on tests. Fable 5.2 scored 69.9 on Frontier SWE. It hit 89.3 on OSWorld 2.1, a computer-control test, against 79.5 for Astra. It also scored 74.2 on Humanity’s Last Exam. It didn’t win everything. Carbon edged it on Automation Bench (53 to 51), and Mythos 5.2 barely beat it on Terminal Bench Science.
I think the real takeaway is how far an idea can go now. With a short prompt and a clear scope, one person can build a sellable game or a full desktop app.
👉 Watch the full video to see the team’s live Tumble Beans match, the Tokyo flyover and the tower carrying that drill!