That’s on cost obv, the other issue w/ GPT-4 is rate limiting /reliability, which if you use any of the above you’re still relying on AN endpoint. But we went from like a theoretical “oh one day” to seeing the actual economics coalesce around what a pretty good model will cost.