Show HN: We benchmarked 18 LLMs on OCR (7k+ calls) – cheaper models often win

Category: devtools

Tags: ocr, benchmark, llm, evaluation, typescript

Score: 7.0/10 (Innovation: 6, Technical: 7, Documentation: 8, Utility: 7)

OCR Mini-bench is a reproducible benchmark for comparing LLM-based OCR extraction quality, cost, and reliability on business documents like invoices and receipts. It's interesting because it provides a structured, multi-run comparison framework with checkpointing and a frontend, challenging the assumption that more expensive models perform best for OCR tasks.

Target audience: backend devs, data engineers, devops

Repository: https://www.arbitrhq.ai/leaderboards/ · TypeScript · MIT · 1 stars

View on Hacker News