There are two rankings on this page. The first is the benchmark. We gave eight systems the same 72 real documents, asked each for the same fields, checked every answer by hand, and ranked them on accuracy and nothing else. You can rerun it yourself from the published data. The second is our vendor scorecard, where we weigh the five dedicated products on everything else you'd care about if you were buying: whether it just works, what it costs, and what you get with it. The benchmark tells you which system reads documents best. The scorecard tells you which product you'd actually want.
How many fields each system got right, averaged over every document. One row per system, at its best setting. Five of them are dedicated extraction platforms; the other three are frontier AI models called directly, which is what you'd get if you built it yourself. There's no judgment in this table.
Share of fields extracted correctly, averaged per document.
The five dedicated extraction products, scored on everything a buyer weighs. Accuracy is the biggest single factor, but it's only a third of the weight. The rest is our own scoring of each vendor's published pricing, product, and policies. The weights are fixed, and every number has a reason written next to it on that vendor's review. Sorted by the weighted total.
| # | System | Accuracy35% | Resilience15% | Price15% | Getting started10% | API and integration10% | Data control7.5% | Compliance7.5% | Weightedout of 10 |
|---|---|---|---|---|---|---|---|---|---|
| 1 | DocuPipe | 9.7 | 10.0 | 6.0 | 10.0 | 9.0 | 9.0 | 10.0 | 9.1 |
| 2 | Reducto | 8.9 | 6.0 | 7.0 | 7.0 | 7.0 | 7.0 | 7.0 | 7.5 |
| 3 |  Extend | 8.0 | 5.0 | 7.0 | 9.0 | 9.0 | 7.0 | 5.0 | 7.3 |
| 4 |  Unstructured | 6.8 | 4.0 | 8.0 | 6.0 | 6.0 | 8.0 | 8.0 | 6.6 |
| 5 |  Pulse AI | 7.1 | 3.0 | 5.0 | 7.0 | 8.0 | 6.0 | 8.0 | 6.2 |
Click a vendor to read the reasoning behind each score. The accuracy column is the benchmark result on a 0 to 10 scale. Every other column is our judgment, based on what the vendor publishes about pricing, product and policies. The three AI models called directly aren't on this card, because they aren't document products and most of these columns don't apply to them. Their reviews still score them the same way: Claude Sonnet 5, GPT-5.5, Gemini 3.5 Flash.
The leaderboard shows one setting per system. Here's the full list, with each product's standard and premium settings next to each other, so you can see what the upgrade actually buys you.
| System | Setting | Accuracy |
|---|---|---|
| DocuPipe | High | 97.02% |
| DocuPipe | Standard | 96.14% |
 Claude Sonnet 5 | Called directly | 91.73% |
| Reducto | Deep Extract | 89.38% |
| Reducto | Standard | 81.11% |
| Extend | Default | 80.28% |
 GPT-5.5 | Called directly | 76.48% |
 Gemini 3.5 Flash | Called directly | 72.98% |
| Pulse AI | Top tier, extended reasoning | 70.95% |
| Unstructured | Auto partitioning | 67.67% |
The same results, cut by the document types that usually decide a purchase. Each slice comes straight from the published document metadata, so you can rebuild any of these numbers from the evidence page.
Hebrew and Arabic layouts, where column order and text direction have to be resolved before a single character is read.
Dense Chinese and Japanese scans, including vertical text and microfilmed sources.
Catalogs, directories, statements and registers with hundreds of entries to keep track of, where most systems either stop early or quietly return a fraction.
