Summary
High-to-medium effort cut slowed Claude; Anthropic restored high default on Apr 7.
A bug cleared prior ‘thinking’ every turn, causing forgetting and repetition; fixed Apr 10.
A 25/100-word system prompt to cut verbosity hurt coding quality; reverted Apr 20.
If you’ve been using Claude Code over the last month or so, you may have noticed that the agent’s quality degraded somewhat. It’s easy to chalk it up to your imagination, but this time, it’s not you; it’s Claude Code. Anthropic has published a blog post detailing three areas where Claude Code had degraded performance since March, and how it fixed each one.
Related
Anthropic cut Claude Code from new Pro subscriptions, calling it an “A/B test”
It’s no longer shown as available for Pro subscribers on the payment plan page.
Claude Code suffered three issues over the last month
It was a rough period for the AI agent
Over on the Anthropic website, the company published a blog post titled “An update on recent Claude Code quality reports.” In it, the company reveals three areas that affected Claude’s performance and caused a notable decline in its abilities. The good news is, since April 20th, all three pain points have been fixed, so Claude Code should be back to normal now.
The first issue began on March 4th, when Anthropic noticed that Claude Code’s default reasoning effort was causing large delays, to the point where people assumed the app had frozen. To remedy it, they swapped the effort level from “high” to “medium,” which limited Claude Code’s capabilities. Anthropic reverted the change on April 7th after users told the company that they’d prefer to have “high” as the norm and reduce the effort value when needed; if they want Claude Code to respond faster, they could always reduce the effort themselves.
The second one appeared on March 26th on Sonnet 4.6 and Opus 4.6. Anthropic tweaked the system to remove Claude’s older thinking if a chat sat idle for over an hour, meaning responses used fewer tokens when the session was resumed. Unfortunately, a bug caused this to happen after every turn, meaning Claude kept forgetting things and repeated phrases a lot. This got fixed on April 10th.
Finally, on April 16th, Anthropic says that it “added a system prompt instruction to reduce verbosity.” Specifically, the prompt said this:
“Length limits: keep text between tool calls to ≤25 words. Keep final responses to ≤100 words unless the task requires more detail.
The change didn’t mix well with other prompting tweaks, causing it to “hurt coding quality,” as Anthropic puts it. The problem got fixed on April 20th.
Related
I set up Claude Code the way its creator does, and the difference is night and day
Who better to learn from than the person who made it?