Anthropic reverts three system changes that degraded Claude Code performance in March and April
Anthropic confirmed three separate system changes in March and April degraded Claude Code, Claude Agent SDK, and Claude Cowork performance. The company reduced default reasoning effort from high to medium on March 4, introduced a caching bug on March 26 that cleared session data with every turn, and added restrictive word limits on April 16 that caused a 3% performance drop.
Anthropic Reverts Three System Changes That Degraded Claude Code Performance
Anthropic confirmed that Claude Code users who reported declining quality in March and April were correct. The company identified three distinct system changes that degraded performance across Claude Code, Claude Agent SDK, and Claude Cowork. The Claude API was not affected.
Three Separate Issues
Reasoning effort reduction (March 4): Anthropic changed Claude Code's default reasoning effort level from high to medium to reduce latency from extended thinking periods. According to the company, "This was the wrong tradeoff." The change was reverted April 7. The current Claude Code build v2.1.118 now defaults to "xhigh" reasoning effort on Sonnet 4.6.
Caching bug (March 26): Engineers introduced a bug while attempting to clear old output tokens for idle users after one hour. Instead of clearing cached thinking sessions only after idle periods, the bug cleared cached session data with every prompt-response turn. This made Claude "forgetful and repetitive." The issue was fixed April 10 for Sonnet 4.6 and Opus 4.6.
Response length limits (April 16): Anthropic added new system prompt instructions to reduce verbosity: "Length limits: keep text between tool calls to ≤25 words. Keep final responses to ≤100 words unless the task requires more detail." Internal testing before deployment suggested the change was safe, but post-deployment ablation tests revealed a 3% performance drop for both Opus 4.6 and 4.7. The system prompt change was reverted April 20.
Company Response
Anthropic emphasized it did not intentionally degrade its models. The company is implementing additional measures including more internal testing for Claude Code builds, improvements to its Code Review tool, better evaluation of system prompt changes, and a new @ClaudeDevs X account for detailed product communications.
The company reset usage levels for all customers following the issues. Head of growth Amol Avasare separately committed to more direct communication after previously addressing an unannounced A/B test through social media.
What This Means
This incident exposes the fragility of complex AI systems where seemingly minor adjustments cascade into measurable quality degradation. The 3% performance drop from word count restrictions reveals how difficult it is to predict the impact of system prompt changes, even with internal testing. More concerning is that Anthropic's initial testing failed to catch the caching bug and the performance impact of length restrictions—suggesting their pre-deployment evaluation framework needs strengthening. The month-long window between introducing issues and fully reverting them indicates a gap in real-time monitoring systems.
Related Articles
Anthropic Report: Claude Was Used to Target US Navy Ships, Build Missiles, and Track Uyghurs
Anthropic's latest threat intelligence report documents five cases where state and non-state actors used Claude for military targeting, weapons development, mass surveillance, and repression. The findings include an Iran-linked operation targeting US naval forces and a Mali-based system capable of monitoring 25 million phones.
Anthropic Threat Report: Claude Used for Missile Software, Mass Surveillance, and Systematic Theft by Chinese AI Labs
Anthropic's latest threat intelligence report covers December 2025 through August 2026, documenting Claude's misuse in espionage, weapons development, and nationwide surveillance operations. The report also details how seven Chinese AI labs ran covert networks—some routing their own customers' requests through Claude—to extract training data at industrial scale.
Analysis: Claude 'Fable 5.1' Drops Em Dashes and Hedging Language, Answers Grow 30% Longer
A new Arena.ai analysis of tens of thousands of Text Arena outputs shows Claude 'Fable 5.1' has shifted its writing style significantly from Fable 5 — using fewer em dashes, less hedging language, and producing 30% longer responses. The codenamed models appear to be unreleased Anthropic checkpoints being tested anonymously on LMArena.
Anthropic Joins Google in Watermarking AI-Generated Text, Reviving Debate Over Output Quality
Anthropic announced on August 11 that all future Claude models will embed an invisible watermark in generated text, following Google's lead with SynthID-Text. The move is partly driven by the EU AI Act, which mandates watermarking for AI models released after August 2, 2026, though researchers remain split on whether the technique degrades output quality.
Comments
Loading...