Models like Deepseek is the only reason we are able to categorize and measure quality of thousands of MCP servers (https://glama.ai/blog/2026-04-03-tool-definition-quality-sco...). That's billions of tokens – an expense that would be otherwise very hard to swallow.