They have several levers for demand destruction. From Anthropic's POV, I suspect this is least to worst bad
- reducing the surface area of "acceptable use" (e.g., blocking third-party tools OpenClaw)
- tighter usage limits and more subscription tiers
- increasing existing subscription prices
- moving to usage based model completely
- taking away compute from training next gen models (future demand destruction)
Reduce quality? Like quantizing models or context cache