Heavy Claude users who upgraded to Opus 4.7 are reporting a noticeable quality drop compared to Opus 4.6, with the most common complaint centered on what one high-volume user calls the "meta" problem - the model spends too much time talking about itself rather than doing the task.
The criticism comes from someone who logged 17 consecutive weeks of maxing out Claude's usage limits on the Max 20x plan, Anthropic's highest-tier subscription. That's not a casual user with vague impressions. The specific complaint: Opus 4.7 "can't stop being meta" - a pattern where the model becomes preoccupied with its own reasoning process, buries outputs in unnecessary caveats, or narrates what it's doing instead of just doing it. The user cited technical research and academic work as primary use cases where this behavior actively gets in the way.
This kind of regression is a known failure mode after model updates where safety or measured quality metrics don't align with what real users need. A model tuned to appear more careful can paradoxically become less useful - better at hedge-writing, worse at clean execution. Opus 4.6 apparently got that balance right in ways 4.7 does not.
The awkward part for Anthropic: Opus 4.6 remains accessible to Max subscribers, which means the current flagship update sits alongside a version some users actively prefer. When experienced users on paid plans are routing around your newest model, that's a signal worth taking seriously.
Anthropstored has not publicly addressed the regression reports.