Why Claude Code Got Worse — and How Anthropic Fixed It: The Full Postmortem (2026)

If you spent March and April 2026 wondering whether Claude Code was getting dumber, you weren't imagining it. Developers across Reddit, X, and Hacker News had been reporting the same thing for weeks: Claude felt less capable, more repetitive, weirdly terse. Some called it AI shrinkflation. Others quietly switched back to older workflows.

On April 23, 2026, Anthropic made it official. They published a postmortem identifying three specific engineering changes that degraded Claude Code's quality — and confirmed all three had been fixed, with subscriber usage limits reset as compensation.

Here's the complete breakdown: what broke, why it broke, and what it means for your workflow going forward.

What Users Were Actually Experiencing

Before diving into the root causes, it's worth cataloguing the symptoms developers reported. These weren't vague impressions — they were consistent patterns across thousands of users:

Reduced reasoning quality on complex tasks. Claude would arrive at simpler, shallower solutions for problems that previously got deep, multi-step analysis. Tasks that required sustained reasoning — refactoring large codebases, debugging race conditions, architecting systems — felt noticeably worse. Memory loss within sessions. Claude would forget context it had already processed earlier in the same conversation. It repeated questions, ignored prior tool call results, and produced inconsistent output across turns. Users described it as "talking to a goldfish." Bizarrely short responses. Responses that should have been detailed — explanations of architecture decisions, breakdowns of test failures, code comments — were cut off at oddly short lengths. Some users got one-line replies to questions that warranted paragraphs. The API wasn't affected. Importantly, developers using the raw Claude API reported none of these issues. The degradation was specific to Claude Code's managed environment — which turned out to be the key clue.

The 3 Root Causes Anthropic Identified

Anthropic's postmortem traces the degradation to three separate changes, each shipped independently, each with a plausible rationale that turned out to be wrong in practice.

1. Reasoning Effort Downgraded to Medium (March 4)

The first change was the most impactful. On March 4, Anthropic changed Claude Code's default reasoning effort from high to medium.

The stated goal was latency. When Claude's extended thinking process runs at full effort, the UI can appear "frozen" for several seconds — a jarring experience when you're waiting for a quick answer. The team believed that defaulting to medium effort would make the product feel snappier without a meaningful quality tradeoff.

They were wrong. Users immediately noticed the difference on tasks that required sustained, multi-step reasoning. The complaints started accumulating within days. Anthropic ultimately reverted this change on April 7, after enough users made clear they'd rather wait an extra few seconds for a smarter answer than get a faster but shallower one.

The key lesson Anthropic acknowledged: default to intelligence, not speed. Users can opt into faster, lighter responses for simple tasks — but downgrading the default without asking is the wrong tradeoff.

2. A Cache Bug That Wiped Short-Term Memory Every Turn (March 26)

The second issue was a subtle but devastating bug, shipped on March 26 as part of a caching optimization.

The intended behavior: when a Claude Code session has been idle for over an hour, prune the "thinking" traces from the cache to save memory and reduce costs. A reasonable optimization — stale context from an old session isn't useful.

The actual behavior: a logic error caused the cache to be cleared on every subsequent turn, not just after an idle timeout. So after Claude finished its first response in a session, every following turn started with a blank slate. The model lost its short-term memory with every message.

This explained the "goldfish" behavior developers had been reporting. Claude would re-ask for information you'd already provided, ignore prior decisions that had been made in the conversation, and produce answers that contradicted what it had said two turns earlier.

The bug was fixed on April 10 for both Sonnet 4.6 and Opus 4.6.

3. A Verbosity System Prompt That Killed Coding Quality (April 16)

The third change arrived on April 16 alongside the Opus 4.7 launch. As part of efforts to make Claude Code less verbose — a real complaint from users who wanted concise outputs — Anthropic added length constraints to the system prompt:

Keep text between tool calls to ≤25 words
Keep final responses to ≤100 words

On paper, this sounds like a reasonable UX improvement. In practice, it was catastrophic for coding tasks.

Anthropic's own ablation testing later revealed a 3% drop in coding quality evaluations for both Opus 4.6 and Opus 4.7 after this change. For coding benchmarks, 3% is a meaningful regression — especially when the constraint was applied uniformly regardless of task complexity.

A one-line explanation of why a regex is failing has different verbosity requirements than a detailed walk-through of why a distributed lock is causing a deadlock. The blanket word limits crushed quality on exactly the tasks Claude Code users care most about.

This change was reverted on April 20.

The Fix: Version 2.1.116

All three issues were resolved by April 20, 2026 in Claude Code version v2.1.116:

Change	Date Introduced	Date Fixed	Impact
Reasoning effort: high → medium	March 4	April 7 (reverted)	Complex task quality ↓
Cache clearing bug	March 26	April 10	Session memory loss
Verbosity system prompt	April 16	April 20 (reverted)	Coding eval quality ↓3%

On April 23, Anthropic reset usage limits for all Claude Code subscribers as compensation — acknowledging that the degraded experience was below what users should expect.

The API remained unaffected throughout. If you were building on the Claude API directly (via the Anthropic SDK or REST endpoints), your applications never saw these issues.

What This Means If You're a Claude Code Subscriber

If you were affected, here's the practical summary:

You should be back to pre-March quality as of April 20. All three changes have been reverted or fixed. The reasoning effort is back to high by default, the cache clears correctly (only after idle timeout), and the verbosity constraints are gone. Check your version. If you're on v2.1.116 or later, you have the fixes. Claude Code updates automatically for most users, but if you're on a managed enterprise deployment, confirm with your admin. Your usage limits were reset. Anthropic reset account limits for all subscribers as of April 23. If you burned through your usage during the degraded period — running more tasks to compensate for lower quality output — that usage has been restored. The API is your escape hatch. This incident is a good reminder that the raw Claude API is more stable than the managed Claude Code environment for production-critical workloads. API users saw none of these issues. If you're building serious agentic workflows, consider routing through the API directly for more control over model behavior and system prompts.

Lessons for Developers Building on Claude

Beyond the fix itself, this incident carries some useful signals for anyone building products on top of Claude:

1. Reasoning effort is a first-class parameter. The fact that dropping from high to medium effort caused this much user backlash confirms that reasoning effort isn't just a latency knob — it's a quality dial. When you're building with the API, be explicit about your thinking budget. Don't let it default to whatever the system sets. 2. System prompt changes are silent breaking changes. Anthropic changed a system prompt and caused a 3% coding quality regression — without any versioned release that developers could pin to. If you're building agents or applications on top of managed products, you have limited visibility into system-level prompt changes. Document your expected behavior and run evals regularly. 3. Managed vs. raw API is a real architectural choice. Claude Code's managed environment provides convenience (tools, sandboxing, UI) but introduces a new layer of risk: Anthropic can change system prompts, reasoning defaults, or caching behavior without notice. Raw API usage gives you more control at the cost of more configuration. 4. User feedback loops matter at this scale. Anthropic caught and fixed all three issues primarily because users complained loudly and consistently. The March 4 reasoning change was reverted within five weeks, the cache bug within two weeks, the verbosity prompt within four days. The feedback → fix loop is working, but it requires developers to report degradation clearly with reproducible examples.

How to Get the Best Out of Claude Code Right Now

With the fixes in place, here's how to maximize Claude Code's performance in May 2026:

Use Claude Opus 4.7 for complex reasoning tasks — it's back to full reasoning effort and benefits from the reverted verbosity constraint
Set explicit context windows — don't rely on implicit session memory; for long tasks, re-inject key context at major checkpoints
Run your own evals — if you're using Claude for consistent code generation tasks, maintain a small eval suite so you catch future regressions before they compound
Prefer API for production agents — for mission-critical workflows, build against the API with explicit model and thinking parameters rather than the managed environment

Key Takeaways

Claude Code's quality drop from March–April 2026 was real, not imagined
Three engineering changes caused it: reasoning effort downgrade (March 4), a cache-clearing bug (March 26), and a verbosity system prompt (April 16)
All three were fixed by April 20 in v2.1.116; usage limits reset April 23
The raw API was unaffected throughout
The incident highlights the importance of pinning model behavior, running evals, and understanding the tradeoffs between managed products and raw API access

Next Steps

If this incident made you want a deeper understanding of how to build reliably on Claude — from choosing models to writing system prompts to configuring reasoning budgets — our Claude API beginner's guide covers the fundamentals from scratch.

Preparing for the Claude Certified Architect (CCA) exam? Our CCA exam guide walks through everything you need to know, and our CCA practice test bank gives you 200+ exam-style questions to test your knowledge — including questions on model behavior, API configuration, and agentic architecture.

Sources: The Register, Fortune, Implicator AI, TechBriefly