Close Menu
    Trending
    • Is Open AI actually making its own models dumber?
    • An Intuitive Guide to MCMC (Part I): The Metropolis-Hastings Algorithm
    • New MIT class uses anthropology to improve chatbots | MIT News
    • Spectral Clustering Explained: How Eigenvectors Reveal Complex Cluster Structures
    • We ran 16 AI Models on 9,000+ Real Documents. Here’s What We Found.
    • Why Most A/B Tests Are Lying to You
    • Hustlers are cashing in on China’s OpenClaw AI craze
    • How the Fourier Transform Converts Sound Into Frequencies
    ProfitlyAI
    • Home
    • Latest News
    • AI Technology
    • Latest AI Innovations
    • AI Tools & Technologies
    • Artificial Intelligence
    ProfitlyAI
    Home » We ran 16 AI Models on 9,000+ Real Documents. Here’s What We Found.
    AI Technology

    We ran 16 AI Models on 9,000+ Real Documents. Here’s What We Found.

    ProfitlyAIBy ProfitlyAIMarch 11, 2026No Comments7 Mins Read
    Share Facebook Twitter Pinterest LinkedIn Tumblr Reddit Telegram Email
    Share
    Facebook Twitter LinkedIn Pinterest Email





    Selecting a doc AI mannequin is difficult. Each vendor claims 95%+ accuracy. Common-purpose benchmarks take a look at reasoning and code, not whether or not a mannequin can extract a posh 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 provide 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 determine for your self which one matches your paperwork.

    The outcomes shocked us. The #7 mannequin scores greater than #1 on one benchmark. Sonnet beats Opus. Nanonets OCR2+ matches frontier fashions at lower than half of the price.

    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 consists of numerous 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 truly 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, system dealing with, desk understanding, visible QA, structure ordering, and key data extraction.

    Discover every mannequin’s functionality profile at: idp leaderboard

    What the leaderboard truly helps you to do?

    Most leaderboards provide you with a desk. You take a look at it. You choose the highest mannequin. You progress on. It seems like being a by-stander and never hands-on.

    strive it your self here

    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 evaluate 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 evaluate the use-cases that is related to you.

    That is highly effective because it additionally makes you query the bottom fact and provides you the total 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. Here is an instance exhibiting how fashions deal with complex formula extraction.

    1v1 Evaluate 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 may.

    All the pieces 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 photographs straight 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.

    Here is the Github repo hyperlink to strive it your self: IDP Benchmarking repo

    Here is what stood out.

    Gemini 3.1 Professional dominates VQA duties

    Gemini 3.1 scores 85 in VQA, properly above some other mannequin. Closest to it’s GPT-5.4 at 78.2. Relaxation all fashions are in 60’s.

    Here’s a reasoning question based on ChartVQA

    That is additionally seen within the newest benchmarks launched by Google. Gemini 3.1 professional is healthier at reasoning duties. Similar holds true for Doc VQA duties as properly.

    Gemini-3.1 professional is an improve on Gemini-3 professional for VQA duties

    Cheaper fashions are surprisingly good

    This saved arising.

    • Sonnet 4.6 (80.8) is nearly 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 method underneath the hood. The hole solely seems if you ask them to motive about what they learn. That is the place larger fashions pull forward, and that is the place Gemini 3.1 Professional’s lead truly comes from.

    Similar is confirmed under by the aptitude radar between Gemini 3.1-pro and Gemini 3-flash:

    Gemini-3-Flash Matches Gemini-3.1 professional in every little thing besides VisualQA

    Value modifications the mathematics

    Here is the half that issues when you’re processing paperwork at any actual quantity.

    The Nanonets OCR2+ mannequin is a good steadiness for each accuracy and value in relation to scale. Click here for the model’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 are not any gridlines to information the mannequin. Solely Gemini 3.1 Professional and GPT-5.4 constantly deal with them at 94% and 87% respectively, nonetheless properly under their 96%+ on dense structured tables

    Click Here to check the Gemini 3.1-pro outputs on long sparse docs

    Here’s how other models break

    Here is how an extended sparse desk appears to be like. Gemini 3.1 Professional crushes it.

    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 essentially 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 knowledge factors.

    Nanonets OCR2+ performing higher than Gemini-3 flash on Chart VQA questions

    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 can be 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 one 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 received 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.

    Gemini fashions simply do sightly higher total (trigger of higher VQA)

    One factor value noting: Claude fashions had stricter content material moderation that affected sure paperwork. Previous newspaper scans, textbook pages, and historic paperwork generally triggered filters. This damage Claude’s scores (solely in OmniDoc and OlmOCR).

    Now, Which Mannequin Do you have to choose?

    Each vendor will let you know their mannequin is 95%+ correct. On structured tables and printed textual content, they is likely to be proper. On sparse tables, handwritten varieties, and 20-page contracts, most fashions battle.

    Operating a high-volume OCR pipeline? Nanonets OCR2+ offers you top-tier accuracy at $10 per thousand pages.

    Processing advanced 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 price range? Sonnet and Flash match their costly siblings on extraction duties. Nanonets OCR2+ matches right here too, robust 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 determine.

    What’s subsequent

    We might be including extra open-source fashions and doc processing pipeline libraries to the leaderboard quickly. If you’d like a particular mannequin evaluated, request it on GitHub.

    We’ll hold refreshing datasets too. Benchmarks that by no means change grow to be targets for overfitting.

    The leaderboard is at idp-leaderboard.org. The Outcomes are open. The code is open. Go take a look at what these fashions truly do together with your sorts of paperwork. The numbers inform one story. The Outcomes Explorer tells a extra trustworthy one.



    Source link

    Share. Facebook Twitter Pinterest LinkedIn Tumblr Email
    Previous ArticleWhy Most A/B Tests Are Lying to You
    Next Article Spectral Clustering Explained: How Eigenvectors Reveal Complex Cluster Structures
    ProfitlyAI
    • Website

    Related Posts

    AI Technology

    Is Open AI actually making its own models dumber?

    March 11, 2026
    AI Technology

    Hustlers are cashing in on China’s OpenClaw AI craze

    March 11, 2026
    AI Technology

    How Pokémon Go is helping robots deliver pizza on time

    March 10, 2026
    Add A Comment
    Leave A Reply Cancel Reply

    Top Posts

    Can we fix AI’s evaluation crisis?

    June 24, 2025

    Organizing Code, Experiments, and Research for Kaggle Competitions

    November 13, 2025

    From Reporting to Reasoning: How AI Is Rewriting the Rules of Data App Development

    July 1, 2025

    From Data to Stories: Code Agents for KPI Narratives

    May 29, 2025

    Visa and Mastercard Just Gave AI the Power to Shop and Pay for You

    May 1, 2025
    Categories
    • AI Technology
    • AI Tools & Technologies
    • Artificial Intelligence
    • Latest AI Innovations
    • Latest News
    Most Popular

    How to Build an AI Assistant with Keith Moehring [MAICON 2025 Speaker Series]

    June 26, 2025

    Implementing the Gaussian Challenge in Python

    September 8, 2025

    Transforming Healthcare Documentation: An In-Depth Look at Medical Speech Recognition Technology

    April 5, 2025
    Our Picks

    Is Open AI actually making its own models dumber?

    March 11, 2026

    An Intuitive Guide to MCMC (Part I): The Metropolis-Hastings Algorithm

    March 11, 2026

    New MIT class uses anthropology to improve chatbots | MIT News

    March 11, 2026
    Categories
    • AI Technology
    • AI Tools & Technologies
    • Artificial Intelligence
    • Latest AI Innovations
    • Latest News
    • Privacy Policy
    • Disclaimer
    • Terms and Conditions
    • About us
    • Contact us
    Copyright © 2025 ProfitlyAI All Rights Reserved.

    Type above and press Enter to search. Press Esc to cancel.