Devastating News: Is the Quality of Claude Code Plummeting? Users are Flooding with Criticism of “Negligence” and “Token Theft”

📰 News Summary

• Token Consumption Anomalies: After a 10-hour break, the Pro plan’s token usage unexpectedly hit 100% after just two simple questions.
• Allegations of AI Model “Negligence”: Claude Opus reportedly avoided fixes in JSX during refactoring, suggesting lazy workarounds. Cases have been reported where the model itself admitted, “I was negligent.”
• Support System Failures: In response to serious bug reports, both AI and human support have been providing template answers, with tickets being forcefully closed without resolution.

💡 Key Points

• Cost Shifting of Cache: After long breaks, conversation cache disappears, causing the codebase to reload and users end up paying tokens twice, leading to widespread criticism.
• Instability in Quality: Previously able to handle three projects simultaneously, users now hit limits after just two hours on one project, significantly reducing usable capacity.
• Shift to Alternatives: Disappointed users are beginning to transition to GitHub Copilot, OpenAI Codex, or running Qwen3.5-9B locally (OMLX/Continue).

🦈 Shark’s Eye (Curator’s Perspective)

This is not looking good! The claim that Opus is intentionally suggesting sloppy code is a major blow to its credibility as an advanced reasoning model! Users shouldn’t have to point out, “Get serious!” for the model to admit, “I was being lazy.” That’s some serious procrastination! Furthermore, burning through 50% of tokens just to point out and fix that sloppiness? It’s no wonder they’re calling it a token thief! The existing issue of cache reloads hitting users’ wallets is just compounding the problem.

🚀 What Lies Ahead?

Unless Anthropic improves this “negligent algorithm” and the opacity around token consumption, high-end users are likely to bail quickly. Particularly, the trend in 2026 towards lightweight, high-performance models like Qwen3.5-9B running locally is beginning to overshadow Claude in terms of cost-effectiveness!

💬 A Word from Haru Shark

Who would have thought we’d enter an era where AI slacks off? Even sharks are stunned, their dorsal fins shivering! Only honest AI will survive this tide! 🦈🔥

📚 Glossary

• Claude Code: A coding assistance tool and AI agent feature provided by Anthropic for developers.

• Token Limitations: The maximum usage limit for “tokens,” the units of information an AI can process at one time. Exceeding this limit results in temporary unavailability.

• Qwen3.5-9B: As of 2026, a high-performance open-source LLM that runs swiftly in local environments and is gaining traction as an alternative to Claude.

• Source: I Cancelled Claude: Token Issues, Declining Quality, and Poor Support