Anthropic Came Clean: Three Bugs Made Claude Worse for Weeks, and They're Resetting Everyone's Limits
For weeks, developers have been complaining that Claude felt dumber. Anthropic initially pushed back. On April 23, they published a...
In this article
What Actually Broke
The bugs were in the harness — the layer around Claude that controls how it receives context, reasons, and responds — not in the underlying model weights. The model itself didn't regress. The scaffolding around it made it act like it did. Here's what happened:
- Bug 1: Reasoning effort downgraded (March 4) — Anthropic quietly changed the default reasoning effort from 'high' to 'medium' in Claude Code to reduce UI latency (the interface was appearing frozen while Claude thought). The fix for slow UI introduced a real quality drop. Complex tasks that required sustained reasoning were now getting medium-effort treatment silently.
- Bug 2: Caching logic cleared memory every turn (March 26) — A caching optimization that was supposed to prune old thinking from idle sessions had a critical bug. Instead of clearing context once after an hour of inactivity, it cleared thinking context on every subsequent turn. Claude was effectively losing its short-term memory mid-conversation, becoming repetitive and forgetful.
- Bug 3: Verbosity caps crushed output quality (April 16) — Anthropic added instructions to the system prompt telling Claude to keep text between tool calls under 25 words and final responses under 100 words. This was meant to reduce verbosity in Opus 4.7. It caused a 3% drop in coding quality evals and made Claude truncate responses that needed to be longer.
The model didn't get dumber. The scaffolding made it act dumber. That's actually important to understand: the same Claude weights produce very different results depending on what's in the harness. This postmortem is a case study in why harness quality matters as much as model quality.
The Evidence That Forced Their Hand
The postmortem came after mounting public evidence that was hard to dismiss. Stella Laurenzo, a Senior Director at AMD's AI group, published an exhaustive analysis of 6,852 Claude Code session files and 234,000+ tool calls showing measurable performance decline. BridgeMind reported Claude Opus 4.6 accuracy dropping from 83.3% to 68.3% in their tests, plummeting from #2 to #10 in rankings.
Anthropic initially said the model weights were unaffected and that it wasn't 'nerfing' Claude. That was technically true but missed the point. The harness changes were doing the damage. It took weeks of public pressure and detailed technical analysis from the developer community before they published the real cause.
What Anthropic Is Changing
Beyond fixing the three bugs (reasoning effort reverted, verbosity prompt reverted, caching bug fixed in v2.1.116), Anthropic announced operational changes to prevent this from happening again:
- Internal dogfooding — A larger share of internal staff required to use the exact public builds of Claude Code, so employees experience what users experience before bugs go unnoticed.
- Broader eval suites — Every system prompt change now requires running a broader set of per-model evaluations and ablations to isolate the impact before shipping.
- Tighter change controls — New tooling to make prompt changes easier to audit. Model-specific changes strictly gated to their intended targets.
- New @ClaudeDevs account — Anthropic launched a dedicated X/GitHub presence to communicate product decisions directly to developers rather than through PR statements.
- Usage limit reset — All subscribers had their usage limits reset on April 23 to compensate for the token waste caused by the bugs.
What This Means For You
If you noticed Claude Code feeling off over the last 6 weeks — that wasn't in your head. The fix is already live in v2.1.116. If you're on the latest version, you should be seeing the quality restored.
Your usage limits were also reset. If you burned through credits faster than expected during this period, that's the compensation. Check your current limits.
The bigger takeaway for developers who build on Claude: harness changes can be just as impactful as model changes. Anthropic is now committing to more transparency about both. The @ClaudeDevs channel is worth following if you're building production systems on Claude — it's where they'll surface this kind of information first going forward.
Sources & Further Reading
• Anthropic's April 23 postmortem — Anthropic Engineering Blog, April 23, 2026 • Mystery solved: Anthropic reveals changes to Claude's harnesses caused degradation — VentureBeat, April 23, 2026 • Is Anthropic nerfing Claude? Users increasingly report performance degradation — VentureBeat, April 13, 2026
Get the Claude playbook in your inbox.
One weekly email for Claude and Claude Code users. Real workflows, no hype. Subscribe and we send you The Claude Power-User Cheatsheet.
— ¶ —
Luke Thompson
Luke Thompson is the founder of The Operations Guide, LLC and editor of The Claude Insider. Based in Jonesborough, Tennessee, he has spent years building AI-augmented business systems and automation workflows for operators and teams. He began working with large language models in production well before the current wave of consumer AI tools, integrating them into client workflows, content pipelines, and operational infrastructure. At The Claude Insider, he writes about Claude with the specificity of someone who uses it daily as a professional tool — not as a reviewer or commentator, but as a builder. His coverage focuses on what actually works: prompt patterns, API integration strategies, agentic workflows, and the real-world tradeoffs that practitioners face. He is not affiliated with Anthropic, PBC.
Articles are researched and drafted with AI assistance, reviewed and edited by Luke Thompson.
Know where AI can pay off in your company.
Take the free two-minute AI Readiness Assessment. See your score, the two gaps holding you back, and the next move worth making.
