When using Claude Sonnet 4.5 via BYOK on AWS Bedrock, I noticed that I spent over $100 in a single day with only moderate usage. After some investigation, I discovered that the default configuration does not enable prompt caching.
Since it would require minimal effort on Cursor’s end to enable prompt caching by default for AWS Bedrock connections, I strongly suggest implementing this improvement. It would make a big difference for users like us who rely on company AWS credits and would like to continue using Cursor efficiently.
For reference, Claude Code via Bedrock already uses prompt caching by default, so it would be great for consistency if Cursor did the same.
I’ve also noticed huge spend on AWS Bedrock because of the lack of caching. Would really appreciate a fix, and it should only require adding a few headers anyway.
Has this been implemented? I am still seeing high costs on AWS Bedrock, could be an issue on our side but want to rule out that its not missing implementation.
Cursor supports prompt caching for requests made to AWS Bedrock. To view request-level details about cache reads and writes, simply hover over the token count on https://cursor.com/dashboard/usage.
If you want to confirm this from the AWS side instead of the Cursor dashboard: CloudWatch, namespace AWS/Bedrock, metrics CacheReadInputTokenCount and CacheWriteInputTokenCount, by ModelId. If writes are big and reads stay near zero, caching is on but missing, and you pay the cache-write premium every turn, which costs more than no caching. Usual causes in agent loops: something dynamic before the checkpoint, the tool list changing between turns, or one turn adding more than 21 blocks (lots of parallel tool calls) so Bedrock’s lookback can’t find the old entry. I made a small free CLI, cachecanary, that reads Bedrock invocation logs and prints hit rate and dollars lost to misses (cachecanary logs <folder>), if you have invocation logging on.