Claude Opus 4.6 fuels nerfing fears

The take is plausible, but not proven: frontier models can feel worse over time for reasons that have nothing to do with the base weights, including routing changes, safety filtering, load-based fallbacks, or scaffold updates.

• –Community complaints line up with the broader “model drift” concern, but anecdotes alone are weak evidence without fixed prompts, fixed accounts, and version logging.
• –Trackers like MarginLab and AIStupidLevel are useful because they make degradation visible over time, not just at launch.
• –The weak spot is adversarial adaptation: once a benchmark matters, providers can optimize for it or quietly whitelist test accounts, which turns the metric into theater.
• –The right answer is a rotating, hidden, and independently operated eval suite, ideally paired with user-side regression tests on real workloads.
• –For developers, the practical lesson is to benchmark your own use case continuously instead of trusting launch-day impressions or leaderboard snapshots alone.