Scanning images with an LLM burns tokens fast. Running OCR first and feeding the text to the model instead drops the cost to near-zero — here's the math and a working demo.
Read the full article