If I’m reading it correctly - Cursor handles prompt cache details on the backend and it’s a non-configurable 5m cache time for Anthropic models. I’m curious why there isn’t an option for 1hr. It’s more expensive for sure, but for long running jobs it can be a big difference maker.
Less of a request and more of a question, what am I missing here?
Hey @Mujahid_Maqsood, thanks for the question.
We actually ran a controlled test of the 1 hour cache across real agent usage. Anthropic charges 2x the base input price for 1 hour cache writes versus 1.25x for 5-minute writes, and across the board, that extra write cost slightly outweighed the savings from fewer cache misses, so the average request ended up a bit more expensive rather than cheaper.
One thing that makes the 5 minute window less painful than it sounds is that Anthropic refreshes the cache lifetime on every hit at no extra cost. During a long running agent turn, every tool call and model step is a hit, so the cache stays warm for as long as the agent is actively working. The window only really matters when a conversation sits idle for more than 5 minutes between messages.
Individual usage patterns can differ from the average, so we keep an eye on them, but right now not in our plans to change.
Thank you for the explanation. This cost is ultimately what the users pay. Why not allow them to experience that / make the call for themselves? I could totally be oblivious to technical challenges or work that would be needed to make this happen so perhaps it is not worth it, but I am curious.