Vendor terms and billing surprises are the new platform risk

Anthropic stopped Claude Code subscriptions from working with OpenClaw, offering a one-time credit instead. Commenters say that is the capacity they paid for, call it a money grab, and one complains of hitting rate limits within 1:45 on afternoons. In a related thread, people compare running local models for daily coding: Qwen 3.6 27B described as roughly Haiku 4.5 level, DeepSeek V4 Flash on two RTX Pro 6000s, and others finding Gemma 4 on an M4 too slow. An AWS billing glitch sent hobby accounts alerts for hundreds of millions to billions of dollars.

The pattern: the cost and the rules of your AI and cloud stack can change under you overnight. The reaction is to look for a way out, whether that is OpenRouter, a local model, or at least spend caps you control.

Local models are not there yet for most people. The replies are full of speed and tooling caveats, so the pressure is real but the escape hatch is partial.