I think for me it's more the possibility of the ideas here being incorporated in other Opensource projects (vllm, llama.cpp, or in the architecture choices of the models themselves)

Deepseek manages to be ridiculously cheap and efficient with prompt caching, i spent 550M~ (1.9b input cache miss, 1.7b output) tokens and it cost me 5~ bucks, that would be ~150 bucks on Opus 5.5, if i ran that workload in parallel that would be $50 vs $1500, e.i. i wouldn't want to subagent heavily with anthropic models.

now imagine Kimi K4 or GLM 5.5 with those cost savings