Query optimization

I entered this challenge on beyondrelational. With the sample data of about 300 rows my solution ran in 0 ms so I thought I had something, but alas the load testing showed it took about 8000ms. Looking at the winners, my solution is almost identical to Stefan's solution below except mine has a replace and I do not do the "pseudo recursion"

-- tc46_v2
-- stefan_G

-- uses pseudo-recursion to avoid executing the patindex more than once for each string
-- handle the case with leading ']' separately
-- use datalength() instead of len() to handle trailing spaces properly

;with 
cte1 as (
    select string, patindex(case when substring(string,1,1)=']' then '%[ -\^-ÿ]%' else '%[^'+substring(string,1,1)+']%' end, string collate latin1_general_bin) as i
    from tc46
    union all
    select * from cte1 where i<0
)
select 
    String, 
    case when i=0 then substring(string,1,1) else substring(string,i,2147483647) end as Result
from cte1
order by datalength(string), string collate latin1_general_bin

How can mine be 8 seconds slower than this, can anyone shed light on what the pseudo-recursion is doing? i will not be <0 so it will only return the anchor part, how is this processed differently than a generic cte? Even without it, shouldn't the patindex only be executed once for each row during the table scan? I am floored at the discrepancy of 8 seconds. Further it shows mine having 4000 writes but Stefan's only 41, something doesn't seem right for how close our solutions are