Everyone's sharing the game clips. The thing in this launch that'll actually show up in a brand's numbers got almost no attention, and it has nothing to do with games.
Anthropic put out a new model called Claude Fable 5 on Monday, and by Tuesday my feed was wall-to-wall people screaming at it. Someone asked it for a Minecraft clone and got a playable browser version in 55 minutes, with caves, biomes and a day-night cycle. The clip cleared 890,000 views before lunch. An investor announced that 3D worldbuilding was "solved."
It is genuinely a leap. The real proof isn't the games, it's that Stripe used it to run a migration across a 50-million-line codebase in a single day, work that would've taken a team months.
But a voxel game in a browser is impressive partly because it's a clean problem. No customers, no payment processor melting down at 2am, no ten-year-old CMS nobody dares touch. Almost everything we get paid to do for clients is the messy opposite of that. So I'd hold off before reading a Minecraft clip as a forecast for your own stack. We've already had clients see one of these videos and ask why their AI tool still needs hand-holding. The honest answer: their tool lives in the real world and the demo doesn't.
I won't pretend we've benchmarked Fable 5 yet. It's two days old. What I'll say is that the number we always watch isn't how many things a model nails, it's what the misses cost. One that builds nine apps and silently breaks the tenth isn't a bargain. And it's pricey: a reviewer testing it live found it burns credits at roughly double the rate of the older model, which still won on value for everyday work.
Here's the bit nobody's posting about. Fable 5 is the best model going for reading images, able to pull exact numbers out of detailed charts and rebuild a working app from a single screenshot. It reads charts, tables and diagrams buried inside PDFs, the ugly stuff businesses drown in: competitor pricing screenshots, the invoice your supplier photographed instead of exporting, a dashboard you can't get data out of.
For the brands we work with, "reads messy images properly" is worth far more than "builds a game." It's the difference between automation that needs someone to type everything up first and automation that just looks at the mess and gets on with it. I'd bet on that being the part of this launch with a real claim on your next quarter, and against most of the people currently shouting about 3D.
One catch worth knowing: to ship something this strong publicly, Anthropic wrapped it in safety filters that quietly send some requests to a weaker model, firing in under 5% of sessions. For normal marketing work you'll never notice. The Ferrari just ships with a limiter.
We'll run it on something real soon, and I'll tell you what broke.