Claude’s answers have become incredibly verbose and full of tics, and the community has a name for it: Claudism. This is not a taste problem. Anthropic never calls Sonnet 5, Opus 5, and Fable 5 a family, but the condition seems rather hereditary. Claude is wordier than it was a generation ago, the extra words make the work worse, and Anthropic’s advice is to just prompt better.

Sonnet 5 lost every trace of the laconicism I had noticed in Sonnet 4. Opus 5 made the pattern plain: the response is less a response than a progress of thought. I ask a question and receive the model thinking about it, with the answer somewhere inside. A shorter reply that simply gave the answer would be more useful. Fable 5 did more overall work, and despite its documented heavier cost, it is the restrained one of the three.

The numbers agree. Arena analyzed tens of thousands of high-reasoning Text Arena responses from Opus and Fable, versions 4.5 through 5. The average Opus reply grew from 158 words to 510, sentences ran 58 percent longer, em dashes appeared 2.3 times as often, and “load-bearing” doubled. Fable 5 sat in between at 316. Artificial Analysis agrees: Opus 5 at max effort burned 140 million output tokens on its Intelligence Index against a 94 million median, the second most verbose rating the site gives. Claude against Claude, the family became more talkative.

Output tokens are the meter on every plan, per million on the API and by weekly limit on subscriptions, so a reply three times as long costs roughly three times as much. Every word it pads into a reply is billed twice, once as output and again as input on my next prompt. The current models all write similarly: circle around a point, then suddenly they act as if they hit the jackpot, only to circle around it further down, burying the true conclusion somewhere in the middle. Digging it out is work the model has created when its sole purpose is to reduce or automate it. Output that buries its answer is lower quality output.

None of this is news to Anthropic. Its prompting guide for Fable 5.1 says the new model uses fewer stock phrases and less unexplained jargon, and it supplies a paragraph defining “mannered prose”. That addresses the tics. It does not address the length. The same section admits sentences run longer with fewer paragraph breaks, and Artificial Analysis clocked Fable 5.1 at 190 million output tokens on its index against 130 million for Fable 5, both at max effort. Claude Code got an opt-in Concise style in August. Every proposed fix lives on my side of the prompt.

The models overview states the company’s position plainly: “If you prefer more concise responses, adjust your prompts to guide the model toward the desired output length.” That does not help. At least the community is showing its dissent with satires: “load-bearing” and how Opus 5 still can’t do one job right. A model that needs a paragraph of instruction before it answers briefly is not a concise model ill-handled by bad prompts. It is a verbose model. Until Anthropic ships a Claude that answers as briefly as the last generation did, every extra word is a defect users pay for on the meter and again in their own time.