Persistent 0% prompt cache hits on GPT-5.5 with Auckland NZ Cloudflare 520s complicating every workaround
OpenAI Developer Community
June 17, 2026
A bit of an update: I was able to seemingly resolve the Cloudflare 520s by bouncing my traffic off a foreign VPS.
This has allowed me to achieve stable API responses, even when using foreground synchronous mode.
However, I still have not managed to achieve stable prompt cache hits, even when routing my context windows through OpenRouter, which is particularly surprising.
My top suspicion is that this is perhaps related to some change in how system prompts are handled. I’m preparing a test framework to pressure test this assumption.
Discussion in the ATmosphere