I’m trying to understand OpenAI prompt caching before using it in our cost estimates. Our support assistant sends the same instructions and tool definitions on most requests, but the cached-token count varies more than I expected.

The customer’s message changes each time. We also include a request timestamp near the beginning of the prompt, and our application sometimes changes the order of the tool definitions when building a request.

What should I keep constant for a useful comparison?

• Does a changing timestamp near the start affect caching for the instructions that follow it?
• Does tool-definition order matter even when the available tools are identical?
• How should I separate prompt differences from timing or model-specific caching requirements?

I’d like to run a small, repeatable test using the reported cached input tokens, rather than guessing from response speed. Which request details and usage fields would you log to explain the differences?