OCR Mini-bench finds budget LLM wins

This is useful less because it crowns one model and more because it attacks lazy model selection with repeatable cost data.

• –The benchmark shows standard document OCR is often a model-fit problem, not a frontier-model problem
• –Cost-per-success is the right framing for extraction pipelines because failed calls still hit the bill
• –The dataset is narrow but practical: invoices, receipts, logistics documents, and ground-truth JSON labels
• –Open-sourcing the framework makes this a template for teams to build their own regression sets instead of trusting generic evals
• –The main caveat is scope: premium models may still matter for messy edge cases, handwriting, long-tail formats, or domain-specific reasoning