tested 5 models on the same 5 scanned invoice packs tonight. invoice + delivery order + receiving slip. typed and handwritten. scored field by field against a checked answer key.
seconds per pack:
claude sonnet 5.5: 11
gpt-6-luna: 13
gpt-6-astra: 19
gpt-6.1-sol (out today): 31
gpt-6-sol medium: 69
gpt-6-sol high: 103
zero wrong values on all of them except luna. it read a tick mark through a quantity as someone changing the number by hand.
more reasoning bought nothing. same values at 69s and 103s.
newest model wasnt the fastest. and it hit "model at capacity" on launch day.
picked astra since it runs on my chatgpt plan. sonnet is the backup.