Thariq (trq212) has faced this before:
Jan 2026: "we absolutely don't change which models we route to or 'nerf' the models"
Apr 2026: "we don't degrade our models to better serve demand"
Aug 2026: "We sometimes test API serving configs... evals confirm this doesn't affect model performance."
Anthropic's Apr 23 postmortem later detailed harness changes (effort defaults, caching bug, verbosity limits) that passed tests yet degraded quality.
显示更多
