Back to Home

Degradation of Claude Code: reasoning 3.5 times less

AMD AI Director analyzed 6852 Claude Code sessions, revealing drop in reasoning depth by 3.5 times and reduction in file reads before edits. API costs grew 80 times due to errors. Anthropic recommends manual effort settings.

Claude Code dulled: AMD report on 6852 sessions
Advertisement 728x90

# Claude Code Degradation: Thinking Reduced 3.5x Based on 6852 Sessions

Stella Laurenzo, Director of the AI group at AMD, analyzed 6852 Claude Code sessions from January–March 2026. Total volume: 234,760 tool calls and 18,000 thinking blocks. Conclusion: the model struggles with complex engineering tasks. Thinking depth fell from 2200 characters in February to 600 in March—correlated with the redact-thinking-2026-02-12 flag, which Anthropic introduced in early March.

Changes in Model Behavior

After flag activation, the share of hidden thinking rose from 1.5% to 100% in a week. This led to fewer file reads before edits: from 6.6 to 2 per edit. The model now makes changes to files without previewing them first, causing:

  • Broken comments.
  • Violations of agreements from CLAUDE.md.
  • Pseudo-fixes that don't solve the problems.

AMD's script for detecting phrases like “continue?”, “known limitation,” or “this wasn't my edit” triggered 173 times after March 8. Previously—zero triggers.

Google AdInline article slot

Scaling and Cost Growth

The AMD team increased parallel agents from 1–3 to 5–10 across 10 projects. API requests surged from 1498 in February to 119,341 in March—80x with similar user prompt volumes. Degradation ramped up token usage: more errors, repetitions, and failed iterations.

Fun fact: the report was generated by Claude Opus 4.6 based on logs. In its self-analysis, the model noted the reads/edits ratio drop from 6.6 to 2.0 and 173 stop attempts, but didn't explain the quality decline.

Anthropic's Response

Claude Code lead Boris Cherny denied any thinking cuts: the redact-thinking flag only hides them in the UI. Recommendations:

Google AdInline article slot
  • Manually activate /effort high or /effort max.
  • For enterprise clients—default high effort levels in tests.

Rumors say AMD temporarily switched to an alternative tool.

Methodology for Assessing Depth

Direct access to thinking is blocked by the flag, so Laurenzo used correlations with other log fields. Data from 6852 sessions confirms the trend: the decline coincided with the flag rollout. AMD senior engineers unanimously deemed Claude unsuitable for engineering work.

Key Takeaways

  • Thinking depth reduced 3.5x (2200 → 600 characters).
  • File reads before edits: 6.6 → 2.
  • 173 script triggers for refusals after March 8.
  • API requests ×80 with team scaling.
  • Anthropic denies degradation, suggests manual effort tweaks.

— Editorial Team

Google AdInline article slot
Advertisement 728x90

Read Next