Claude max-effort regresses, tool use duplicates emails

This smells more like an orchestration regression than a model-brain collapse: users feel Claude got worse because effort settings, routing, and tool execution all live on the same trust boundary. Once max-effort stops feeling dependable, every automation becomes a liability.

• –Anthropic still says Opus 4.6 is the strongest option for the deepest reasoning, and its Sonnet 4.6 launch claims fewer false success claims and less duplication, so the video hits a core promise.
• –The March 23, 2026 status incident was narrowly about elevated errors on Claude.ai, with the API unaffected, so it does not fully explain duplicate-email behavior in downstream workflows.
• –Tool-use regressions are worse than chat regressions because they create real-world side effects, not just a bad answer.
• –If effort defaults or model routing changed under the hood, Anthropic needs to expose that clearly; otherwise users will keep calling any quality dip a nerf.
• –The clip resonates because Claude's value prop is exactly careful reasoning plus agentic action, and the host says both slipped at once.