Abstract
- Top-to-medium effort minimize slowed Claude; Anthropic restored top default on Apr 7.
- A malicious program cleared prior ‘pondering’ each flip, inflicting forgetting and repetition; fastened Apr 10.
- A 25/100-word gadget instructed to chop verbosity harm coding high quality; reverted Apr 20.
If you happen to’ve been the use of Claude Code during the last month or so, you’ll have spotted that the agent’s high quality degraded relatively. It is simple to chalk it as much as your creativeness, however this time, it isn’t you; it is Claude Code. Anthropic has revealed a weblog put up detailing 3 spaces the place Claude Code had degraded efficiency since March, and the way it fastened each and every one.
Anthropic minimize Claude Code from new Professional subscriptions, calling it an “A/B check”
It is now not proven as to be had for Professional subscribers at the fee plan web page.
Claude Code suffered 3 problems during the last month
It was once a coarse duration for the AI agent
Over at the Anthropic website online, the corporate revealed a weblog put up titled “An replace on fresh Claude Code high quality reviews.” In it, the corporate unearths 3 spaces that affected Claude’s efficiency and led to a notable decline in its talents. The excellent news is, since April twentieth, all 3 ache issues had been fastened, so Claude Code will have to be again to commonplace now.
The primary factor started on March 4th, when Anthropic spotted that Claude Code’s default reasoning effort was once inflicting huge delays, to the purpose the place other folks assumed the app had frozen. To treatment it, they swapped the trouble stage from “top” to “medium,” which restricted Claude Code’s features. Anthropic reverted the trade on April seventh after customers advised the corporate that they would wish to have “top” because the norm and scale back the trouble price when wanted; if they would like Claude Code to reply quicker, they might all the time scale back the trouble themselves.
The second seemed on March twenty sixth on Sonnet 4.6 and Opus 4.6. Anthropic tweaked the gadget to take away Claude’s older pondering if a talk sat idle for over an hour, that means responses used fewer tokens when the consultation was once resumed. Sadly, a malicious program led to this to occur after each flip, that means Claude saved forgetting issues and repeated words so much. This were given fastened on April tenth.
In spite of everything, on April sixteenth, Anthropic says that it “added a gadget instructed instruction to scale back verbosity.” In particular, the instructed mentioned this:
“Period limits: stay textual content between device calls to ≤25 phrases. Stay ultimate responses to ≤100 phrases until the duty calls for extra element.
The trade did not combine smartly with different prompting tweaks, inflicting it to “harm coding high quality,” as Anthropic places it. The issue were given fastened on April twentieth.
I arrange Claude Code the way in which its author does, and the adaptation is evening and day
Who higher to be told from than the one who made it?
