The worst part is not even the cap, it's that coding sessions are too bursty for provider-side budgeting to feel sane. If you still have to think in Anthropic windows and token math, the bottleneck never really went away. That's a big part of why we're building rig.ai around local inference.