Which AI reads your documents best? Test it on your own files. Free.
Send 20 of your real documents. We run them through 4 AI models and tell you which is most accurate, what it costs, and where it fails. Reply within 48 hours.
No sales call required. NDA on request. Files deleted after the report.
Same score. 80× price gap.
We tested 9 service configurations on 6 synthetic one-page invoices and delivery notes, with 10 fields checked on each. 7 of 9 got every field right. Their estimated cost differed 80×.
Field accuracy vs. cost per page
October 2026 pilot · 6 synthetic documents
- 1OpenAI gpt-4.1-mini
- 2Google gemini-2.5-flash
- 3Mistral OCR
- 4Anthropic claude-sonnet-4-5
- 5LlamaParse · Cost Effective
- 6AWS Textract · AnalyzeExpense *
- 7LlamaParse · Agentic
- 8LlamaParse · Agentic Plus
- 9Google Document AI · Invoice Parser *
Sorted by estimated cost. Cost bar uses a log scale.
| Service / model | All fields correct | Field accuracy | Median time | USD / 1,000 pages | Cost |
|---|---|---|---|---|---|
| OpenAI gpt-4.1-mini | 6/6 | 100% | 1.87 s | $0.93 | |
| Google gemini-2.5-flash | 6/6 | 100% | 2.40 s | $1.16 | |
| Mistral OCR | 6/6 | 100% | 3.02 s | $5.00 | |
| Anthropic claude-sonnet-4-5 | 6/6 | 100% | 2.86 s | $7.76 | |
| LlamaParse · Cost Effective | 6/6 | 100% | 10.30 s | $10.00 | |
| AWS Textract · AnalyzeExpense | 0/6 | 80%* | 3.52 s | $10.00 | |
| LlamaParse · AgenticRun twice, same score | 6/6 | 100% | 9.81 s | $31.25 | |
| LlamaParse · Agentic Plus | 6/6 | 100% | 12.26 s | $75.00 | |
| Google Document AI · Invoice Parser | 0/6 | 70%* | 1.76 s | $100.00 |
* AWS Textract and Google Document AI don't return a document-type field, so they lose that field on every document. Without it, their accuracy is about 89% and 78%. Not scored: Azure Document Intelligence (not run) and LlamaParse Turbo (request failures).
What this pilot does not show
- 6 clean synthetic documents can't predict results on your scans, photos or handwriting.
- It is too small to rank vendors. Several scored 100% because the test was easy.
- Costs are estimates from public prices, not invoices. Models change often.
What it does show
- A clean demo makes every vendor look perfect.
- The price gap is real: 80× for the same score on the same files.
- The only test that matters is on your own documents.
Fixed scope. Clear output.
- 01
Accuracy per field
Invoice number, date, total, supplier — which model got each one right.
- 02
Cost at your volume
Price per 1,000 pages for each model, and per month for your volume.
- 03
Where it fails
The files each model got most wrong, with what went wrong.
- 04
A clear recommendation
Which model to use — or whether none is reliable enough yet.
How it works
- Step 1
Tell us the basics
Document type, fields you need, rough monthly volume.
- Step 2
Send 20 processed files
Documents you've already entered in your system, plus an export of those values.
- Step 3
We run and check
4 models, scored against your values. Disagreements checked by hand.
- Step 4
One-page report in 48 h
Optional 20-minute call to walk through it.
Why your system export?
Accuracy needs correct answers to compare against. You already have them: the values your team typed in. No export? We label the 20 files ourselves and say so in the report.
Your data
- NDA signed before upload, if you want one.
- We tell you where each model processes data.
- Files deleted within 7 days of the report.
- Never used to train any model by us.
The only test that matters is on your own documents.
Send 20 documents