Claude Confirms Downgrade: 3 Bugs Exposed

Post LinkedIn

⚛️Read original on 量子位

#model-degradation #conversation-bugs #quota-resetclaudeclaude anthropic

💡Claude dumber confirmed—3 bugs, quotas reset. Test apps before production fails!

⚡ 30-Second TL;DR

What Changed

Model intelligence downgrade officially confirmed

Why It Matters

This degradation impacts reliability for long-form AI applications, forcing developers to reassess Claude dependencies. Quota reset offers temporary relief but signals ongoing stability issues.

What To Do Next

Check your Anthropic Console for reset Claude quotas and test long-conversation performance immediately.

Who should care:Developers & AI Engineers

Key Points

•Model intelligence downgrade officially confirmed
•Performance degrades in prolonged chats
•Three bugs fully exposed
•All usage quotas reset for users

🧠 Deep Insight

Web-grounded analysis with 4 cited sources.

🔑 Enhanced Key Takeaways

•The performance degradation was localized to Anthropic's agentic tools—Claude Code, Claude Agent SDK, and Claude Cowork—while the core Claude API remained unaffected.
•The three identified bugs included a forced reduction in default reasoning effort (March 4), a faulty session-idle memory-clearing mechanism that caused repetitive behavior (March 26), and an overly restrictive system prompt limiting verbosity (April 16).
•Anthropic attributed the difficulty in identifying these issues to the fact that each bug affected different traffic slices on different schedules, creating an aggregate appearance of random, inconsistent model degradation.

📊 Competitor Analysis▸ Show

Feature	Anthropic (Claude)	OpenAI (GPT-4/o)	Google (Gemini)
Agentic Tooling	Claude Code / Cowork	Operator / Assistants API	Gemini Code Assist
Reasoning Control	Explicit Effort Levels	System Instructions	Adaptive / Auto-scaling
Context Handling	1M+ Token Context	128k - 2M (Model dependent)	1M - 2M Context
Recent Stability	Public post-mortem on bugs	Periodic user reports of 'laziness'	Generally stable / high throughput

🛠️ Technical Deep Dive

•Reasoning Effort: Anthropic uses an 'effort' parameter in the Messages API to trade off latency for reasoning depth; the default was mistakenly lowered from 'high' to 'medium' on March 4.
•Memory Management: A bug in the session-idle logic caused the system to clear reasoning history on every turn rather than once after an hour of inactivity, leading to 'short-term memory' loss.
•System Prompting: A specific instruction was injected to limit text between tool calls to ≤25 words and final responses to 100 words, which negatively impacted coding quality.
•Resolution: All three issues were resolved by April 20, 2026, in version v2.1.116 of the affected tools.

🔮 Future ImplicationsAI analysis grounded in cited sources

Anthropic will implement stricter internal 'dogfooding' requirements.

The company announced that a larger share of internal staff must now use public builds of Claude Code to ensure parity between internal and external user experiences.

Future system prompt changes will undergo mandatory 'ablations'.

To prevent similar regressions, Anthropic is introducing a broader suite of per-model evaluations to isolate the impact of specific system prompt instructions.

⏳ Timeline

2026-02

Launch of Claude Opus 4.6 with 1M token context and new effort controls.

2026-03

Introduction of bugs affecting reasoning effort (March 4) and session memory (March 26).

2026-04

Reversion of reasoning effort change (April 7) and memory bug fix (April 10).

2026-04

Implementation of verbosity prompt (April 16) and subsequent resolution of all three bugs (April 20).

2026-04

Official post-mortem published and usage quotas reset for all subscribers (April 23).

📎 Sources (4)

Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.

⚛️Read original article on 量子位

📰

Weekly AI Recap

Read this week's curated digest of top AI events →

👉Related Updates

Same topic

Explore #model-degradation

Same product