Thanks for reaching out!
Both storage, read and write costs are linear by the record size. Therefore, a decent approximation for 3072-dim costs would be 2x the values of the 1536-dim vectors. So, for example:
- Upserting a single 3072-dim vector with 1000 bytes of metadata would use 14 WUs
- Querying a 1M-records namespace with 1000 bytes of metadata would use 36 RUs
- Storing 10M records at 3072 would use ~132 GB
Does that answer your question? If not, feel free to provide more information about your specific case.