If downloading more than 5% of stored data is so expensive, wouldn't it have been cheaper to upload a file 19 times the size of the stored data (containing /dev/urandom)?
After that, downloading just 5% of total data would have been free.
It still wouldn't have been free. The free download allowance is spread daily across the entire month. That is, you can download 5% of your data per month, and only 0.16% of it per day. For your optimization to work, you'd have to retrieve your data over 30 days for it to be free.
At that point, your monthly fee would be greater than that of a regular S3 bucket, which has fewer barriers to retrieval.
Reminds me of my advice to netflix after peering problems emerged, just push garbage upstream from every client to equalize your peering traffic.