using gpt5 to ocr tables tables and flowcharts is a pain. a sota model from a top AI lab still has trouble reading tables. it is an interesting benchmark for the model, rather than most benchmarks out there where you can game the system. although it is hard to mark that the model did well. the output format is just reconstruct it in a way a doctor can interpret and recreate it, there's no fixed format or language easily verifiable like code. it requires human label and review which is costly. i found some techniques where you use a package first then pass that info to gpt, or have gpt output structured data like json representing table rows and headers, which could work.