Selecting a doc AI mannequin is tough. Each vendor claims 95%+ accuracy. Common-purpose benchmarks check reasoning and code, not whether or not a mannequin can extract a fancy desk from a scanned bill.
So we constructed the Clever Doc Processing (IDP) Leaderboard.
3 open benchmarks. 16+ fashions. 9,000+ actual paperwork. The duties that matter: OCR, desk extraction, key data extraction, visible QA, and lengthy doc understanding.
The purpose is not to present you one quantity and declare a winner. It is to allow you to dig into the specifics. See the place every mannequin is powerful, the place it breaks, and resolve for your self which one matches your paperwork.
The outcomes shocked us. The #7 mannequin scores increased than #1 on one benchmark. Sonnet beats Opus. Nanonets OCR2+ matches frontier fashions at lower than half of the fee.
Why 3 benchmarks?
Each benchmark measures one thing totally different. Use one and also you solely see one dimension. So we used three.
OlmOCR Bench: Are you able to reliably parse a messy web page? Dense LaTeX, degraded scans, tiny-font textual content, multi-column studying order. Fashions that excel at one usually fail at one other. This dataset contains various set of pdfs.
OmniDocBench: Does the mannequin perceive the doc’s construction? Formulation, tables, studying order. Structure comprehension, not simply character recognition.
IDP Core: Are you able to extract what a enterprise really wants? This one is ours. Invoices, handwritten textual content, ChartQA, DocVQA, 20+ web page paperwork, six sorts of tables. The stuff that breaks manufacturing pipelines. These are extra reasoning heavy duties than the opposite two benchmarks.
Every mannequin will get a functionality profile throughout six sub-tasks: textual content extraction, formulation dealing with, desk understanding, visible QA, structure ordering, and key data extraction.
Discover every mannequin’s functionality profile at: idp leaderboard
What the leaderboard really permits you to do?
Most leaderboards offer you a desk. You have a look at it. You choose the highest mannequin. You progress on. It appears like being a by-stander and never hands-on.
We wished one thing extra clear and hands-on than that.
For that we created the Outcomes Explorer that permits you to see precise predictions and examine fashions on actual paperwork. For any doc within the benchmark, you see the bottom fact subsequent to each mannequin’s uncooked output. This makes you see and examine the use-cases that is related to you.
That is highly effective because it additionally makes you query the bottom fact and offers you the complete image of what is going on behind the scenes of every benchmark activity.
You possibly can see precisely the place it hallucinated a desk cell or missed a handwritten phrase. This is an instance exhibiting how fashions deal with complicated formulation extraction.
1v1 Examine places two fashions aspect by aspect throughout all six functionality dimensions.
How did we run it?
We wished anybody to have the ability to run all three benchmarks. So we made setup as near zero as we might.
Every part pulls from HuggingFace. We pre-rendered all PDFs to PNGs and hosted them at shhdwi/olmocr-pre-rendered so you do not want a conversion pipeline. IDP Core embeds photos instantly within the dataset. Nothing to clone your self or unzip.
The runner works with any mannequin that has an API. Failed runs choose up the place they left off.
This is the Github repo hyperlink to attempt it your self: IDP Benchmarking repo
This is what stood out.
Gemini 3.1 Professional dominates VQA duties
Gemini 3.1 scores 85 in VQA, effectively above another mannequin. Closest to it’s GPT-5.4 at 78.2. Relaxation all fashions are in 60’s.
That is additionally seen within the newest benchmarks launched by Google. Gemini 3.1 professional is best at reasoning duties. Identical holds true for Doc VQA duties as effectively.
Cheaper fashions are surprisingly good
This stored arising.
- Sonnet 4.6 (80.8) is pretty much as good as Claude 4.6 (80.3)
- Gemini-3 flash matches Gemini-3 professional and generally even higher (in Omnidoc bench)
This might level to one thing attention-grabbing. Cheaper fashions match costly ones on extraction. Textual content, tables, structure, formulation. They appear to be studying paperwork the identical manner beneath the hood. The hole solely seems while you ask them to purpose about what they learn. That is the place greater fashions pull forward, and that is the place Gemini 3.1 Professional’s lead really comes from.
Identical is confirmed under by the potential radar between Gemini 3.1-pro and Gemini 3-flash:
Price adjustments the mathematics
This is the half that issues should you’re processing paperwork at any actual quantity.
The Nanonets OCR2+ mannequin is a good steadiness for each accuracy and value in the case of scale. Click on right here for the mannequin’s full profile
The place issues nonetheless break!
Sparse, unstructured tables stay the toughest extraction activity.
Most fashions land under 55%. These are tables the place cells are scattered, many are empty, and there aren’t any gridlines to information the mannequin. Solely Gemini 3.1 Professional and GPT-5.4 persistently deal with them at 94% and 87% respectively, nonetheless effectively under their 96%+ on dense structured tables
Click on Right here to verify the Gemini 3.1-pro outputs on lengthy sparse docs
Handwriting OCR hasn’t crossed 76%. One of the best mannequin is Gemini 3.1 Professional at 75.5%. Digital printed OCR is 98%+ for frontier fashions. Handwriting is a basically totally different downside and no mannequin has cracked it.
Chart query answering is unreliable. Nanonets OCR2+ leads at 87%, Claude Sonnet follows at 85%, GPT-5.4 drops to 77%.
The failures are particular: axis values misinterpret by orders of magnitude, the mistaken bar chosen, off-by-one errors on intently spaced information factors.
Handwritten kind extraction hallucinates on clean fields. Each mannequin clusters between 80-84% on this activity. The failure mode is constant: fashions fill in values for fields which are clean on the shape. A reputation, a date, a standing that does not exist within the doc.
Gemini > Claude = OpenAI
The pecking order was settled. Gemini led, Claude adopted, OpenAI trailed. GPT-4.1 scored 70.0. No person was choosing OpenAI for doc work.
For GPT-5.4 Desk extraction went from 73.1 to 94.8. DocVQA went from 42.1% to 91.1%. GPT-5.4 acquired higher at understanding paperwork and reasoning.
The general scores are actually 83.2, 81.0, 80.8. Shut sufficient that the rating issues lower than the form. Claude leads on formulation. GPT-5.4 leads on tables and QA. Gemini leads on OCR and VQA.
One factor value noting: Claude fashions had stricter content material moderation that affected sure paperwork. Outdated newspaper scans, textbook pages, and historic paperwork generally triggered filters. This harm Claude’s scores (solely in OmniDoc and OlmOCR).
Now, Which Mannequin Must you choose?
Each vendor will inform you their mannequin is 95%+ correct. On structured tables and printed textual content, they may be proper. On sparse tables, handwritten types, and 20-page contracts, most fashions wrestle.
Working a high-volume OCR pipeline? Nanonets OCR2+ provides you top-tier accuracy at $10 per thousand pages.
Processing complicated tables or want excessive accuracy on reasoning over paperwork? Gemini 3.1 Professional is well worth the premium at $28/1K pages.
Constructing a easy extraction workflow on a finances? Sonnet and Flash match their costly siblings on extraction duties. Nanonets OCR2+ matches right here too, sturdy accuracy with out the frontier price ticket.
However do not take our phrase for it. The leaderboard has the scores. The Outcomes Explorer has the precise predictions. Decide a activity that matches your workload. Take a look at what they output on actual paperwork. Then resolve.
What’s subsequent
We will probably be including extra open-source fashions and doc processing pipeline libraries to the leaderboard quickly. If you would like a selected mannequin evaluated, request it on GitHub.
We’ll maintain refreshing datasets too. Benchmarks that by no means change change into targets for overfitting.
The leaderboard is at idp-leaderboard.org. The Outcomes are open. The code is open. Go have a look at what these fashions really do along with your sorts of paperwork. The numbers inform one story. The Outcomes Explorer tells a extra sincere one.
