You can look at my solution for clarification of my comments. It isn't that mine is right (because we get to the same solution), but since you asked I will give you critical feedback.
When you see a post by @AdamR, read it.
I went down the same path with counting each letter. Later, I realized the rules of the anagram didn't care about counts. What really mattered was that the alphabetized order of all letters between the two words had to be an exact match: LISTEN & SILENT both become: EILNST
If you solve each word to become the component letters like that, if they match then they are anagrams.
We both could improve our workflows with that in mind.
Please find my script enclosed.
At first i was aiming to use the fuzzy match with the key generation on Whole Field and matching on levenshtein distance but somehow (and i was like a day on it without understanding why to day) it only compared and output a match on the first word, everything else was a match between the same word ( meaning Record 1 with Record 1)... I've tried the Purge and Merde mode (with different IDs on word 1 and 2 to make it looks like it's coming from two diferent sources), I've tried different key generation (Double metaphone) and matching (Jaro distance), it's driving me nuts.
I can post this script that failed but if someone has an explaination that will ease my sleep.