Mystery solved: Anthropic reveals changes to Claude's harnesses and operating instructions likely caused degradation

2026-04-23 · Source: VentureBeat · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Data Science & Analytics · Depth: Intermediate, quick

Summary

Anthropic has addressed user reports of "AI shrinkflation" in its Claude models, which suggested a degradation in reasoning capability, increased hallucinations, and token waste. The company published a technical post-mortem identifying three product-layer changes, not underlying model weight regressions, as the cause. These included a default reasoning effort change from `high` to `medium` for Claude Code on March 4, a caching logic bug introduced on March 26 that cleared thinking history too frequently, and system prompt verbosity limits added on April 16. These issues affected Claude Code CLI, Claude Agent SDK, and Claude Cowork, but not the Claude API. Anthropic has since reverted the reasoning effort and verbosity prompt changes, fixed the caching bug in v2.1.116, and reset subscriber usage limits.

Key takeaway

For AI architects and engineering leads deploying large language models, this incident highlights the critical impact of "harness" or product-layer changes on perceived model performance. Ensure your teams implement robust internal testing, comprehensive evaluation suites for prompt modifications, and strict gating for model-specific adjustments to prevent unintended regressions and maintain user trust. Proactive communication and transparency are key to managing community expectations.

Key insights

Product-layer changes, not model degradation, caused Claude's perceived performance issues.

Principles

• Harness changes impact model behavior.
• Caching logic can degrade model memory.
• Prompt changes alter model quality.

Method

Anthropic identified three specific product-layer changes: default reasoning effort, a caching logic bug, and system prompt verbosity limits. They resolved issues by reverting changes and fixing the bug.

In practice

• Implement internal dogfooding for public builds.
• Run broad evaluation suites for prompt changes.
• Gate model-specific changes strictly.

Topics

• Claude AI Degradation
• Anthropic Post-Mortem
• AI Performance Benchmarks
• System Prompt Engineering
• Caching Logic Bugs

Code references

• anthropics/claude-code

Best for: CTO, VP of Engineering/Data, AI Architect, Machine Learning Engineer, AI Engineer, Director of AI/ML

Related on AIssential

• 🙀 Anthropic accidentally leaked Claude Code's entire source code — Effective AI agent design prioritizes orchestration and memory discipline over raw model intelligence.
• An update on recent Claude Code quality reports — Three distinct changes caused Claude Code's perceived degradation, highlighting the complexity of LLM product management.
• Anthropic Found Out Why AIs Go Insane — AI persona drift causes models to deviate from helpfulness, but activation capping can mitigate this without performance loss.
• Claude Opus 4.7 Leaked, Anthropic Full Stack App, New GPT Model, M2.7, Claude Code Update! AI NEWS — AI model performance can fluctuate due to resource shifts, new releases, or unconfirmed server-side changes.
• Wtf Anthropic — Anthropic's compute miscalculation and opaque policies are alienating users, while OpenAI aggressively captures market share.
• Harnessing Claude’s intelligence — Building with evolving AI requires adapting agent harnesses to leverage new model capabilities and optimize performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by VentureBeat.