Did Anthropic ‘dumb down’ Claude Code? Post-mortem reveals the three bugs that crippled performance

Over the last few weeks, social media has been filled with complaints about Anthropic’s Claude Code feeling slower and dumber, which was evident in the coding quality and the increased memory issues. Anthropic has now confirmed that its AI coding tool was indeed experiencing issues as it released a post-mortem on Thursday, revealing what went wrong with the tool.

In a post on X (formerly Twitter), ClaudeDevs, an account affiliated with Anthropic, wrote, “Over the past month, some of you reported Claude Code’s quality had slipped. We investigated and published a post-mortem on the three issues we found. All are fixed in v2.1.116+ and we’ve reset usage limits for all subscribers.”

“The issues stemmed from Claude Code and the Agent SDK harness, which also impacted Cowork since it runs on the SDK. The models themselves didn’t regress, and the Claude API was not affected,” it added.

Users complain about Claude Code

Forums like Reddit and X have been filled with complaints about Claude Code slowing down and feeling worse for weeks. Users speculated that the decline in quality may be an effort by Anthropic to save money.

“Since last week this has been noticeably slow… like taking minutes to read a simple document… writing the responses line by line and taking its time to do so…” wrote one user on Reddit.

“Claude Code has become so superficial and stupid I can’t trust anything it does anymore,” complained another user.

“I’m trying to figure out if it is nerfing or that it basically shits itself with the 1m context limit when you sandwich different tasks in the same thread. Not sure — feels like a downgrade!” another user complained.

What went wrong with Claude Code?

Anthropic’s post-mortem traced the reports of Claude Code quality issues to three separate updates rolled out between early March and mid-April.

“We’ve traced these reports to three separate changes that affected Claude Code, the Claude Agent SDK, and Claude Cowork. The API was not impacted,” the company wrote in its blog post.

First, Anthropic had lowered the default “reasoning effort” from high to medium in early March to reduce long wait times that made the UI appear frozen for some users. While this improved speed, Anthropic says it also made the model feel less intelligent to users. Anthropic eventually rolled back this change in April after feedback from users, and Opus 4.7 defaults to ‘xhigh’ effort now, while all other models default to ‘high’.

Second, Anthropic had updated the caching optimisation in late March to clear Claude’s older thinking from sessions that had been idle for over an hour.

A caching-related bug introduced later in March caused Claude to repeatedly forget prior context in longer sessions. Instead of clearing older reasoning just once after inactivity, the system kept wiping context on every turn, making the AI appear forgetful, repetitive, and inconsistent in its coding decisions. This issue was fixed in early April.

However, the implementation had a severe bug. Instead of clearing the thinking history once, it wiped the history on every single turn for the rest of the session. This stripped Claude of its memory regarding why it was making certain tool choices, resulting in severe forgetfulness, odd tool choices, and repetitive loops.

“Claude would continue executing, but increasingly without memory of why it had chosen to do what it was doing. This surfaced as the forgetfulness, repetition, and odd tool choices people reported,” Anthropic wrote.

The bug was eventually patched on 10 April.

Third, in preparation for the launch of its newest Opus 4.7 model, Anthropic added a system prompt instruction on 16 April to reduce the model’s verbose nature. The prompt explicitly forced the AI to keep text between tool calls to under 25 words and final responses under 100 words.

By aggressively limiting response length, the update led to a reduction in coding quality, resulting in weaker and less detailed outputs. Anthropic’s tests found that the update had led to a 3% drop in performance for both Opus 4.6 and Opus 4.7 models.