Extend inference costs scale with complexity

The key strategic point is that Extend is not selling a flat OCR utility, it is running a variable cost underwriting business on every page. A clean digital PDF can go through Light Parse at 0.5 credits per page, while harder pages get routed into Performance Parse at 2 credits per page, and extraction layers on more cost again. Auto routing pushes gross margin to depend on what customers send, not just how many pages they send.1 2 3 4

• The routing decision happens at the page level, not just the document level. Extend says Auto mode classifies page difficulty and sends simple pages through the cheap pipeline, while dense tables, handwriting, checkboxes, or messy scans get performance level treatment. That makes a 100 page packet with 10 hard pages much cheaper than 100 equally hard pages.2 4
• The pricing table shows how big the spread is. A 10 page extraction costs 12 credits in Light mode versus 50 credits in Performance mode, and Operator-1 extraction is 5 credits per page before separate parsing charges. This means product mix, parse mode, and document complexity can move unit economics far more than raw page count.1 3 5
• That differs from simpler document APIs from cloud vendors that charge more like standardized page processing. AWS Textract and Google Document AI price core processing per page or per processor call, while Extend exposes a finer ladder of modes and routing logic. The advantage is better margin control on mixed workloads, but only if routing stays accurate.5 6 7

The likely direction is toward more granular orchestration, where Extend keeps adding classifiers, review agents, and specialized modes so expensive model calls are reserved for the small share of pages that truly need them. If that routing gets better over time, Extend can widen the gap between customer price and inference cost without asking customers to manually tune every workflow.2 3 4