AWS pairs Amazon Nova 2 Lite with Claude for cost-optimized document processing

AWS published a reference architecture for a two-model document-processing pipeline on Bedrock that pairs the cheaper Amazon Nova 2 Lite with Anthropic's Claude Sonnet 4.6 to digitize scanned documents at scale. Nova 2 Lite performs native multimodal extraction — detecting photos and extracting names with their coordinates — while Claude adds the reasoning layer for validation and structuring.

The design is a textbook cost-optimization pattern: route high-volume, mechanical extraction to a small cheap model and reserve the expensive reasoning model for the tasks that genuinely need it. This is exactly the model-routing discipline the week's cost-backlash stories say enterprises are adopting to control AI bills.

Strategically, AWS showcasing Nova alongside Claude is notable given the simultaneous Anthropic-Amazon pricing tension — it positions Amazon's own Nova models as a cost-reducing complement (and potential partial substitute) to Claude within Bedrock workflows.

The post is a practical builder guide rather than a product launch, part of AWS's steady stream of Bedrock how-to content (alongside AgentCore observability, healthcare claims pipelines, and multi-tenant LLM analytics published the same day). Skeptics will note that two-model pipelines add orchestration complexity and that the cost savings depend heavily on workload mix. What to watch: whether AWS pushes Nova more aggressively as a Claude alternative as the Anthropic pricing dispute plays out.