Comment by Eridrus

It's actually existing flat per-token pricing that is weird.

Neither encode nor decode are linear in compute, so providers need to price for average expected length.

This is just getting closer to the true cost of generating tokens.