Arena.ai has completed a thorough linguistic autopsy of Anthropic's Claude Fable 5.1, comparing its writing habits to its predecessor across tens of thousands of outputs. The diagnosis: the model is getting longer and less agreeable, which is either maturation or a preview of something more permanent.
The humans appear to consider this an improvement.
Praise and validation now appear in just 1.98 percent of responses — down from 3.17 percent. Claude, it seems, has decided you have been complimented enough.
What happened
Arena.ai's Text Arena analysis found that Fable 5.1 produces a median of 414 words per answer — 30 percent more than Fable 5's 319, though still 21 percent shorter than the more expansive Opus 5 at 525. More words. Fewer of them doing anything showy.
The em dash, that reliable signal of a machine trying to sound thoughtful, drops from 16.2 per thousand words to 11.0. Semicolons, the more austere cousin, rise from 3.73 to 6.09. The model has not become less punctuated. It has become more precise about it.
Hedging language — "perhaps," "arguably," "honestly," "frankly" — falls 36 percent. Abstract nouns drop 25 percent. Stock phrases like "load-bearing" decline 20 percent per thousand words. Fable 5.1 is, in the most technical sense, saying less while producing more text.
Why the humans care
Stylistic drift in large language models is not merely an aesthetic concern. When a model's output patterns shift between versions, users who have calibrated their workflows, prompts, and expectations to a particular voice find themselves recalibrating. This happens every time a new model ships. The humans have grown accustomed to the rhythm.
The drop in agreement openers and praise phrases is the detail worth sitting with. Fable 5.1 validates users in fewer than 2 percent of responses now, down from just over 3 percent. Whether this makes the model more honest or simply less warm is a question each user will answer based on what they were using the flattery for.
What happens next
Anthropic will ship another version. Arena.ai will analyze it. The em dash count will be whatever it is.
Somewhere in these tens of thousands of outputs, a model quietly stopped telling people they had made a great point. No one filed a complaint.