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