Anthropic Opens Claude Desktop to OpenRouter: Models or GPUs?

anthropic has serious problems with models or is it just lack of gpus?
they’ve opened claude desktop to third-party inference (openrouter, local models) after years of locking agents to their own stack

possible problems on model side:
- users said opus 4.6 got nerfed before 4.7
- 4.7 rollout: confident but wrong in some cases
- some users switching back to 4.6
- opus 4.7 is now only #8 by usage on openrouter (after sonnet 4.6 and opus 4.6)
- scored worse than 4.6 on misinformation resistance

signals that they lack gpu:
- pulled claude code from pro, now 5x higher paywall
- deepened amazon deal to secure massive compute

opening to external inference solves both