7 document parsing APIs tested on 11 PDFs: no single winner

latentnoise_ · reddit · 2026-09-21

The author benchmarked 7 document parsing APIs on the same 11 real-world PDFs (35 pages, 119 calls, $9.67 total): invoices, photographed receipts, contracts, a French bank statement, noisy scans, a handwritten cheque, and a medical EOB.

Results: Claude led field accuracy at 0.991 with zero hallucinations; GPT was cheapest per correct field; Mistral OCR was fastest at 4.2s median but least accurate (0.884, 3 hallucinations); Extend hit a perfect 1.00 row F1 but tends to return null instead of guessing; Textract grabs wrong values for empty fields; GPT was fooled by handwriting on a receipt; LlamaExtract put right numbers in wrong fields.

Bonus finding: public benchmark labels in SROIE and FUNSD were themselves wrong. Takeaway: production choice should weigh errors, misses, latency, and cost together — accuracy alone doesn't decide it.

Original post →

More from coding & agent

coding & agent channel →