Mo BitarJune 15, 202621m

I had Fable build several projects for me. I'm disturbed by what I saw.

TL;DR

Fable produced unusually polished one-shot demos: Mo had it build multiple browser games and piano tools, including a Factorio-like AI lab game called Token Mania, an ethereal relic-collecting space game, and a Rust-inspired crafting world with logic gates.
The output felt "beautiful" in a way benchmarks miss: He says the visuals, music, and low-friction coding experience made it feel like "playing inside of Fable's mind," echoing an X post about measuring "the shape of a model's mind" rather than benchmark scores.
The hard part is still the last 20 percent: A bug like weird translucent water in a 3JS demo becomes a nightmare if the model wrote black-box code you do not understand, which is why Mo insists software engineering is not solved.
Expertise still determines how impressed you are: Mo deliberately chose games because he is not a game developer, noting that LLMs tend to wow non-experts and annoy experts who can see the bad writing or unmaintainable code underneath.
Anthropic's capability story looks more credible to him now: After using Fable, he says the company may have been right about Mythos-style gains and accelerating model improvement, even if claims like Matt Shumer's "100 hours with Opus equals 1 hour with Fable" still sound exaggerated.
The economics are the biggest reality check: Mo estimates his demos might have cost $2,000 to $3,000 in API credits, calls the model "completely economically unviable" at current pricing, and argues the flashy $200 subscription era cannot last.

The Breakdown

A Claude variant called Fable felt so uncannily good at one-shot coding that people on X mourned its removal like "their wings had been torn off," but the demos hide the real problem: gorgeous 80 percent outputs still collapse when you need to debug, ship, or pay the bill. Mo Bitar shows several games and music tools Fable built for him, then argues that Anthropic may have been more right than critics wanted to admit about capability gains, even if the economics and maintainability are still brutal.