AWS pairs Nova-2-Lite with Claude to cut document costs

AWS built a document pipeline that routes bulk extraction to Nova-2-Lite and saves Claude for the hard steps, per its machine learning blog.

The split matters for any team running multi-step AI pipelines. Paying frontier prices for high-volume, low-judgment work like OCR buys nothing extra in return.

Nova-2-Lite handles the OCR-grade extraction, the high-volume work that doesn't need judgment calls. Claude only enters for the hard steps, the reasoning phase where judgment is required instead of extraction.

The pattern isn't specific to documents. Any pipeline with a bulk-grunt phase and a reasoning phase can copy the split. Cheap model for volume, frontier model for judgment.

The take

Nova-2-Lite plus Claude reads like a reference architecture other vendors will copy, not a one-off AWS trick. Pipelines still sending every step to one frontier model pay full price for work Nova-2-Lite could do far cheaper. Watch for more vendors publishing their own two-model split under different names.

The write-up sits on AWS's machine learning blog, part of the coverage dated June 30.