For sure they use some caching. The other thing is the computational capacity available - load might change within seconds. In the midst of "thinking" too. The

For sure they use some caching. The other thing is the computational capacity available - load might change within seconds. In the midst of "thinking" too. The currency here is the amount of tokens to be processed and delivered back. The outcome depends on that :)