It's "time to add a word" says Arnold Reinhold, the creator of Diceware, in his blog (3/2014). He advices to use 6 word sentences (or 5 words with one extra character chosen and placed at random) from now on. Reasons given include that he predicted the likelihood of a change in 2014 back in 1995, and that "Today criminal gangs probably have access to more computing power then the NSA"
There is evidence and there are hints (Wikipedia (On the other hand ..) and stackexchange (my comment on Goldberg's answer) that there are (slight?) weaknesses in the Diceware word lists (not in the method). Understanding the he evidence in "Improving the Diceware memorable passphrase generation system" is beyond my math capabilities. I do understand though that the average recovery time of a 4 word sentence can indeed be reduced because ~22 character sentences are the most likely ones. And if the passphrase character count is somehow known, even the exhaustive recovery time may be reduced.
However, Arstechnica mentiones 3/2014 (update) an email received from Gosney in which he would have written "Since there are no tools that currently combine three or more words, we don't really know for sure how much slower it would be compared to his 25 GPU monster cracks."
Are there really no tools and performance measurements available for cracking 3 word or larger Diceware passphrases? I have tried to find measurements and can only find grammar based proof of principle recovery of phrases.
Added in 2 steps after the answers of Thomas Pornin and Arnold Reinhold
Arnold Reinhold acknowledges one word list weakness (Diceware words can run together, “act” and “ion” form "action"). Arnold Reinhold: "I now recommend that users of Diceware put a space character between each word, which completely eliminates this problem." Yet, what is a new formula for calculating entropy of a passphrase when spaces are not used?