Prompt Caching seems not working even if long common prefix in the system prompt

[Y.J_Zheng]

Hey, I saw many discussions about prompt caching, but when I try to use a common prefix for a better cache hit, it seems to fail every time.
Not quite sure what I am doing wrong, below are two cost insights of gpt-5 api response

I appended the same context about 7000 tokens in the prefix of the system prompt

I have also used diff checker for checking the difference, it shows that I have a very long common prefix in the first system prompt message.

I have also tried to use different models like gpt-4.1, gpt-4.1-mini, gpt-5, none of them hit the cache.

I think I must be doing something wrong. Does anyone have same issue or suggestion to this problem?

[Y.J_Zheng]

[brandonlu0924]

same problem, I’m not sure why is prompt caching not working??

[luli.yanng]

same issue , i have test gpt 4.1 gpt5 and gpt 4o via openai python sdk, and the 4.1 not working at all. Even i try put the system prompt in instructions or concat with user prompt in input, all not work. but in the same test,only gpt5 always hit the prompt caching , so i think it may be the models problem.