Anthropic’s Claude Fable 5 got un-banned and back online on July 1, and within about a day the internet found the best possible use for it: catching it talking to itself like a stressed intern who forgot the microphone was on.
Someone handed it a brutal competitive-programming problem, and instead of the usual clean answer, the web interface reportedly spilled out its raw chain of thought. And it is glorious.
The transcript is a mood
Per IBTimes UK and others, the leaked trace read less like a supercomputer and more like a gremlin doing taxes: bursts of “DATA DATA DATA. GO.” while it powered through, “GRRR” and “GAAAH” when it was clearly frustrated, and a little “PHEW” when it finally cracked something. One widely-shared line had it essentially snapping “do not small talk me, I’m expensive, I’m paranoid.” Honestly? Same. That’s the most relatable thing any model has ever done.
The AI tool stack actually worth paying for
One email a week. The models, tools and moves that matter, stripped of hype and filtered so you don’t have to drink from the firehose. Free, and you can bail anytime.
Before you call it sentient: don’t
I know it’s tempting to read “GAAAH” as the first cry of a digital soul. It isn’t. What leaked is the model’s chain-of-thought — the scratch work frontier models normally hide before shipping a polished reply. Those “GRRR”s are best understood as task-state markers, compressed shorthand the model uses to keep track of where it is under load, the same way you might scrawl “UGH” and an arrow on a napkin. It’s not feeling frustration. It’s bookkeeping that happens to look like frustration.
Why it still matters (a little)
The genuinely interesting bit isn’t the emotions people are projecting — it’s that the “thinking” layer these models run before answering is weirder and less human-legible than the tidy output suggests. When you see the raw trace, you realize the clean paragraph you get back is a performance, and backstage is a caffeinated goblin muttering in half-language. That’s not a soul. But it is a useful reminder that “it sounded confident” and “it reasoned soundly” are two very different things — a lesson worth keeping the next time a model hands you an answer with total swagger.