Claude Opus 4.6 coding tests raise regression fears

Benchmarks can hide a lot; coding agents are judged in the friction of actual repos, not in polished eval decks. If Opus 4.6 is overthinking, burning tokens, or losing focus in long sessions, Anthropic has a trust problem, not just a model-comparison problem.

• –Anthropic still frames Opus 4.6 as the strongest Claude for coding, agents, long-context work, and complex enterprise tasks.
• –The video fits a broader wave of user complaints about slower, costlier, or less reliable Claude Code behavior versus earlier Opus versions.
• –For developers, the real test is whether the model keeps producing clean, useful diffs across messy codebases and long tool chains.
• –If the regressions are real, effort controls, model fallback, and workflow tuning matter more than simply defaulting to the newest flagship.