Add multiple models and pricing for each model.

Switch VLM backend to Claude Sonnet 4.6; improve extraction accuracy to 79.9%

Change the model used for the OCR extraction.