For example, size of a transaction including 10 bech32 inputs and 1 output would be about 722 vbytes and 1524 bytes. That's more than 52% reduce in transaction fee.
I found it surprising that segwit to segwit with 10 input and 1 output to have around 52% reduction truly if compared to legacy to legacy. But, I do not think the more reduction in fee is beyond 10 inputs as I calculated for 20 inputs, 80 inputs, even 2000 inputs, while the output remain constant at 1, the results are still around 52% reduction.