Amazon Nova 2 Lite + Claude Sonnet 4.6 cut document processing costs by 66%

Amazon Web Services and Anthropic have published a solution that pairs Amazon Nova 2 Lite with Claude Sonnet 4.6 on Amazon Bedrock to digitize complex scanned documents efficiently. The pipeline tackles the challenge of extracting unstructured information from scanned yearbook pages—detecting portrait photos, reading printed names, and correctly associating each name with the corresponding face. In stage one, Amazon Nova 2 Lite uses native multimodal extraction in a single Converse API call to detect photos (with bounding boxes), extract visible names and their approximate positions, and return page-level metadata. By constraining extraction to names only (rather than full OCR), output tokens are kept to about 1,000 per page instead of an estimated 4,500, minimizing cost. Reasoning level was set to LOW with no accuracy loss, further reducing expenses.

In stage two, Claude Sonnet 4.6 performs spatial reasoning using the positions from Nova to match names to faces, handling layout variability without additional prompt engineering. Running this pipeline against 336 scanned yearbook pages produced 3,122 name-to-face associations, with 93 percent scoring at or above 0.95 confidence. The two-model approach costs roughly two-thirds less per page than a single-model alternative that sends the entire task to one vision-language model. A recent change to Amazon Nova 2 Lite's billing—fixed per-image pricing regardless of resolution—makes per-page costs predictable at scale, a critical advantage for enterprise document processing workflows processing hundreds of thousands of pages.