A 5-word Diceware passphrase gives an entropy of 7776^5 = 3E19 = 19 Bans (or ~64 bits). Apparently monster cracking systems can currently guess passphrases for MD5 at a rate of 180 billion/sec (2E11). See monster system
That system would find my passphrase with an entropy of 19 bans in at most 3E19/2E11=1.5E8 seconds. Or, since 4 month have 1E7 seconds, it would take on average 0.5(1.5E8/1E7) = 0.8E1=8 of 4 month periods, or about 2.5 years. More precise: 0.5(2.8E19/(1.8E11*3.7E7)=2.4 years.
The term on average outrules, I think, a crack estimate of my single Diceware passphrase. Suppose I know the attacker's crack sequence:111111 11112 etc. Then I can calculate how long cracking my passphrase such as "afire akin been snout tress" would take.
If my math is right, it would take about 10x6667x6667x6667x6667=2.0E16 guesses (suppose afire is word 10in the list) which is much less then 0.5* 3E19 for large sets of passphrases when on average starts working.
Can the "If I know the attacker's crack sequence" perhaps be broken and can "on average" be brought back by first randomly sorting a Diceware word list before throwing the dices?