Claude Code users who complained about degraded performance over the past month were right all along. Anthropic’s April 23 engineering postmortem confirmed what thousands of developers suspected: the tool had been quietly sabotaged by three separate engineering mistakes while the company initially suggested users were the problem.

This incident offers critical lessons for any AI engineer building production systems on top of third party tools.

What Actually Happened

Between March 4 and April 20, Anthropic shipped three changes that collectively made Claude Code noticeably worse at its primary job: helping developers write code.

Date	Change	Impact
March 4	Default reasoning effort reduced from “high” to “medium”	Users reported the system felt “less intelligent”
March 26	Caching bug cleared reasoning context on every turn	Claude became forgetful and repetitive
April 16	25-word limit added between tool calls	Coding quality dropped measurably

The first change was a deliberate tradeoff. Anthropic reduced reasoning effort to address UI freezing issues, betting that faster responses would outweigh reduced capability. Users disagreed strongly enough that the company reversed course on April 7.

The second issue was a genuine bug. A prompt caching optimization meant to clear stale reasoning from idle sessions instead cleared it on every turn. This made Claude appear to forget its own context mid-conversation, frustrating users who relied on coherent multi-step interactions.

The third change combined with other prompt modifications to produce a 3% drop in evaluation performance. Anthropic reverted it after four days.

The Trust Problem

What makes this incident particularly concerning isn’t the bugs themselves. Complex systems fail. What matters is how Anthropic initially responded to user complaints.

According to Fortune’s reporting, the company initially blamed users, suggesting they misunderstood the changes. Some users described feeling “gaslit” when their legitimate concerns were dismissed. Enterprise customers reported measuring a 47% drop in code quality while the company maintained nothing was fundamentally wrong.

The detailed engineering postmortem only arrived on April 23, weeks after complaints began. By then, multiple users had cancelled subscriptions and developer trust had eroded significantly.

For AI engineers building on top of these tools, this pattern should feel familiar. When your AI coding assistant stops performing as expected, distinguishing between “user error” and “vendor degradation” becomes critical.

Practical Lessons for AI Engineers

Never Fully Trust Single Tool Dependency

The engineers who weathered this incident best were those with fallback options. When Claude Code started underperforming, having experience with alternative AI coding tools meant work could continue while the primary tool recovered.

This isn’t about abandoning tools at the first sign of trouble. It’s about maintaining enough fluency across options that you aren’t paralyzed when one degrades.

Track Your Own Quality Metrics

Enterprise customers who detected a 47% quality drop did so because they measured it. Most individual developers don’t track AI tool performance systematically, which means degradation can feel like personal productivity decline.

Simple practices help: save before and after versions of AI assisted code, note completion rates on routine tasks, and track how often you accept versus reject suggestions. When metrics shift dramatically, you have evidence that something changed beyond your own skills.

Trust Your Experience Over Vendor Messaging

Experienced developers knew Claude Code was performing worse. They were right. The lesson here isn’t that vendors always lie, but that your direct experience with a tool often provides signal before official acknowledgment.

When something feels off, document it. Share observations with colleagues using the same tools. Community discussion forums often surface patterns faster than vendor support channels.

What This Means for Tool Selection

This incident doesn’t disqualify Claude Code as a serious tool. The company published a thorough postmortem, identified specific causes, and implemented fixes. That transparency, however late, matters.

But it does reinforce why understanding the trade offs between AI coding tools matters more than finding the “best” option. Every tool will have incidents. The question is whether you’ve built enough resilience into your workflow to handle them.

Warning: The combination of Anthropic’s rapid growth and these quality issues has increased speculation about compute constraints. Users reported that the company also suffered outages, introduced usage caps during peak hours, and limited rollout of newer models. Production systems relying heavily on Claude Code should have contingency plans.

The Broader Pattern

This incident fits a broader pattern in agentic AI development. As tools become more capable and autonomous, the surface area for subtle degradation expands. A chatbot that gives slightly worse answers is noticeable. A coding assistant that introduces more bugs or forgets context mid-session can waste hours before you realize what happened.

Anthropic’s postmortem mentioned that Opus 4.7 detected one of the bugs through its own Code Review capability while Opus 4.6 missed it. This suggests an interesting future where AI tools might help audit each other’s quality. Until that becomes reliable, the responsibility falls on engineers to maintain healthy skepticism.

Frequently Asked Questions

Has Anthropic fixed the Claude Code issues?

Yes. All three issues identified in the April 23 postmortem have been resolved. The reasoning effort default was restored on April 7, the caching bug was fixed on April 10, and the verbosity restriction was reverted on April 20.

Should I stop using Claude Code?

Not necessarily. The tool remains highly capable, and Anthropic’s transparent postmortem demonstrates willingness to address problems. However, you should diversify your tooling and track performance metrics to catch future degradation early.

How can I tell if my AI coding tool is degrading?

Track completion acceptance rates, time spent fixing AI generated code, and whether the tool maintains context across longer sessions. Sudden shifts in these metrics suggest tool side changes rather than your own performance.

Recommended Reading

• AI Coding Tools Decision Framework
• AI Coding Errors Troubleshooting Guide
• AI Coding Tools Comparison Guide

Sources

• An update on recent Claude Code quality reports - Anthropic Engineering

To see how production AI engineers handle tool dependencies and build resilient systems, watch the full breakdown on YouTube.

If you want direct guidance on navigating incidents like this while building production AI systems, join the AI Engineering community where members share real time observations, workarounds, and help each other stay productive regardless of which vendor is having a bad week.