Anthropic blames Claude Code defaults for quality drop

This is a strong reminder that “model quality” is not just weights and benchmarks; product-layer defaults can materially change what users experience.

• –Lowering reasoning effort to save latency and usage made the system feel less intelligent, even though the underlying model did not change.
• –The caching bug shows how a small harness mistake can create the exact symptoms users interpret as forgetfulness or regression.
• –The verbosity prompt change is the most interesting one: trimming output can quietly damage coding quality when it interacts with other prompt assumptions.
• –The open-weight/local-model argument gets real support here, not because hosted models are inherently bad, but because control over serving behavior matters when quality is part of your workflow.
• –Anthropic’s transparency is better than most vendors, but the core risk remains: paying customers can still inherit silent quality tradeoffs they did not approve.