How accurate is a $0.75/1,000-pages OCR API? We measured.

Cheap OCR earns instant suspicion: what did they cut? So we built an eval harness and measured our own API the way we'd want a vendor measured — with ground truth that's perfect by construction, degradations included, and the numbers published whether we liked them or not.

METHODOLOGY

We generate synthetic receipts programmatically — merchants, line items, totals, tax — and render them to images. Because we render them ourselves, the ground truth is exact. Half the set gets real-world damage: gaussian blur, rotation up to 3°, salt noise, low contrast. Twenty receipts, scored on three metrics: character error rate (CER) against normalized text, price recall (the fraction of ground-truth amounts that appear in the output — the metric that actually pages you at 2am), and wall-clock latency through the public API.

RESULTS — PENNYOCR, TEXT MODE

The engine is NuExtract3, a 4B vision-language model, served quantized with speculative decoding on commodity GPUs. Markdown mode (tables as HTML, headings, reading order) scores the same on price recall; the eval also caught a repetition-loop failure mode in markdown generation that we shipped a guard for — that's what evals are for.

PRICE PER 1,000 PAGES

Public list prices, first tier, as of August 2026.

HONEST LIMITATIONS

This benchmark is synthetic and receipt-shaped: it measures transcription fidelity under controlled degradation, not every document type. We have not yet run the big clouds through the same harness — that comparison is coming, and the harness plugs into any OpenAI-compatible or REST OCR endpoint precisely so it can. Until then: the honest claim is not "more accurate than X", it's 99.5% measured accuracy at half the price — verify it yourself on the free tier.