I tried using deepseek v4 flash with OpenRouter. It switches between providers too eagerly which resets the cache. Then, each provider begins to rate limit me for providing so many uncached tokens, so it just keeps on switching providers. I'm paying for every token... why rate limit me? It was unusable compared to just using the official Deepseek provider which has a much better cache rate.

I just checked because I was a bit paranoid, but I have a 96.6% cache hit rate for GPT-5.6 Luna and 96.8% for Opus 5.