claude-code - 💡(How to fix) Fix [Bug] Anthropic API: Claude 3.5 Opus thinking effort degradation without disclosure [1 comments, 2 participants]

Bug Description What happened Anthropic silently reduced the thinking effort of Opus 4.6. I have always operated on high-effort settings — this is not a default I overlooked. The quality degraded materially and without warning regardless of that setting. This is not a configuration issue — it is confirmed externally by a public GitHub issue from the AMD Head of AI, and widely reported across developer forums. Financial damage I am hitting my $200/month limit not because I use the product more, but because degraded output quality forces me to iterate 3–5x more per task. I am paying the same price for a fraction of the value. Business impact I have directly referred approximately 500 Team licenses to Anthropic customers. CMI Informatik alone accounts for over 100 licenses. I have invested significant time building an Anthropic-native ecosystem — connectors, skills, agents, workflows. That investment is now at risk due to a reliability failure I had no warning of. I am migrating to model-agnostic tooling (OpenCode) specifically to avoid this single-vendor risk repeating. I will stop recommending Anthropic products in my consulting engagements until reliability and communication standards improve. The core violation Reducing service quality while charging the same price, without disclosure, is — at minimum — an ethical breach and potentially a consumer protection issue under Swiss and EU law. Transparent communication would have been sufficient: "We are experiencing scaling constraints and have temporarily reduced default thinking depth. We are working to restore full quality." That message would have retained my trust. Silence did not. What I'd expect. An answer to this complaint (not from a bot), better and more honest communication. And finally an outlook what is planned on this issue.

Support interaction My initial support contact was handled by an AI agent (Fin) that offered generic workarounds (/model, ultrathink, /feedback) — including suggesting I switch to high effort, which I was already on (always). It did not escalate. It did not acknowledge the confirmed regression. This compounded the problem.

Simon Scheurer Founder, Contem GmbH — Teklens [email protected] account on [email protected]

Environment Info

• Platform: darwin
• Terminal: iTerm.app
• Version: 2.1.110
• Feedback ID: 42ef7c8b-b98c-4f0e-bee9-f203084d50d8