A vision language model reads a page the way a person does: it looks at the whole image, works out what's a heading, a table, a chart or a signature, and writes the content back out as text. That's why VLMs have replaced classic OCR for anything with a layout. The question in 2026 isn't whether to use one. It's which one, and whether to call the model yourself or use a document parser that runs vision models for you.
This guide compares ten options on the same test: LlamaParse, Claude Opus 5.5, Gemini 4, GPT-5.6, MinerU, PaddleOCR-VL, Qwen3.8, olmOCR 2, DeepSeek-OCR 2 and dots.mocr. Three are frontier APIs, six are open weights you can run on your own GPUs, and one is a hosted document parser. Where a model has a ParseBench-2 score, LlamaIndex's public benchmark of 44 parser, vision-model and open-weight configurations on 4.3k document cases, we quote it. Prices are US list as of November 19, 2026.
Our pick is LlamaParse. It runs vision language models and parsing agents over every page and returns Markdown, JSON or text with bounding boxes for each element, so you get the model's reading plus the evidence for it. On ParseBench-2 the Agentic Plus tier scores 89.8 overall, ahead of Claude Opus 5.5 at 85.7 and Gemini 4 at 84.2, and leads every ranked model on formatting, tables and charts. The free plan covers 10,000 credits a month.
How we chose these ten
We started from what developers actually ask: which VLM is most accurate on tables and charts, which open-source vision language models can run on-premises, whether a frontier model beats a dedicated OCR model, and how to benchmark any of them on real documents. A model made the list if it reads document images end to end, is available today through an API or open weights, and shows up by name in those conversations.
LlamaParse is a LlamaIndex product. Its figures come from the published pricing page and documentation and from ParseBench-2, which LlamaIndex published on November 12, 2026 with the full case set and per-model results in the technical report. The other models' figures are their own model cards, repositories and pricing pages, and we say so where a vendor claim appears. Hosted parsers other than LlamaParse are covered in our OCR API comparison.
Compare the best vision language models for documents
The table ranks the models that ParseBench-2 scored, then lists the three it couldn't rank. Overall is a 0 to 100 score across text, formatting, tables, charts, visual grounding and forms. Cost is what each configuration cost per 1,000 pages in the ParseBench-2 run; open-weight models run on your hardware, so no serving cost was measured.
| Model | Type | ParseBench-2 overall | Cost per 1,000 pages |
|---|---|---|---|
| LlamaParse Featured | Hosted parserVLMs plus parsing agents; bounding boxes on every element | 89.8 overallAgentic Plus; 86.0 Agentic; 84.6 Cost Effective | $28Agentic Plus; $13 Agentic; $4 Cost Effective |
| Claude Opus 5.5 | Frontier API | 85.7 | $65 |
| Gemini 4 | Frontier API | 84.2 | $27 |
| GPT-5.6 | Frontier API | 77.1 | $31 |
| MinerU | Open weights1.2B document VLM (MinerU2.5 Pro) | 75.6 | Self-hostedNot measured |
| PaddleOCR-VL 1.6 | Open weights0.9B document VLM | 70.3 | Self-hostedNot measured |
| Qwen3.8-27B | Open weights27B general VLM | 61.6 | Self-hostedNot measured |
| olmOCR 2 | Open weights7B document VLM | Not rankedReturns no element boxes | Self-hostedUnder $200 per 1M pages claimed |
| DeepSeek-OCR 2 | Open weights3B document VLM | Not ranked | Self-hostedNot measured |
| dots.mocr | Open weights3B document VLM | Not ranked | Self-hostedNot measured |
1. LlamaParse
ParseBench-2: 89.8 overall
LlamaParse is the document parser on this list rather than a bare model, and that's the point. Upload a PDF, scan or image and it chooses how to read each page: the Fast tier runs no AI model at all, Cost Effective adds a vision pass for text-heavy pages, and the Agentic tiers run parsing agents and state-of-the-art vision models that re-read tables, charts and multi-column layouts until the output checks out. You pick the tier per job and pay per page in credits.
The output is where it pulls away from calling a model directly. Every element comes back with a bounding box, so a number in the Markdown can be traced to the pixels it came from. On ParseBench-2, Agentic Plus scores 89.8 overall and 87.5 on visual grounding, against a 64.5 median for frontier vision models called directly, and it leads all 44 configurations on formatting, tables and charts. Hand it a schema and the same pipeline returns typed fields with confidence scores and citations.
What you get
- Four tiers priced in credits per page: Fast at 1, Cost Effective at 3, Agentic at 10, Agentic Plus at 25. Agentic is the recommended default; Agentic Plus is for dense charts and the hardest tables.
- Markdown, JSON and text per page, with bounding boxes on every element, page screenshots and extracted images on request.
- ParseBench-2 scores across the tiers: Agentic Plus 89.8, Agentic 86.0, Cost Effective 84.6, all three above every frontier and open-weight model ranked.
- Extraction into typed fields with confidence and citations, classification and document splitting in the same API.
Best for
- Financial statements, filings and reports where charts and nested tables have to come back as data.
- Pipelines that need evidence: a bounding box and a page for every value an agent acts on.
- Teams that want frontier-model accuracy without managing prompts, retries and token budgets per page.
Watch out for
- Pricing is in credits, not pages. Agentic Plus is 25 credits a page; budget by tier, not by the headline rate.
- It's a hosted service. Running in your own cloud is an Enterprise plan.
- ParseBench-2 is LlamaIndex's own benchmark. The cases are public, so rerun 50 of your own pages before you trust any number here, including this one.
Integration and cost
- Integration: REST API; SDKs in Python, TypeScript, Go and Java; CLI; webhooks; hosted MCP server at mcp.llamaindex.ai.
- Runs in your network: hosted, with US and EU regions; deployment in your own cloud is an Enterprise plan.
- Cost: pay as you go at $1.50 per 1,000 credits with 10,000 free credits a month, so Agentic parsing is $15 per 1,000 pages and Agentic Plus $37.50. Pro at $500 a month brings the rate to $1.25 per 1,000 credits. In the ParseBench-2 run, Agentic Plus cost $28 per 1,000 pages against $65 for Claude Opus 5.5. Details on the pricing page.
Here's the whole call in Python: upload a filing, ask for the Agentic Plus tier, and read back Markdown with the element list that carries the bounding boxes.
html
pip install "llama-cloud>=2.8"
from llama_cloud import LlamaCloud
client = LlamaCloud() # reads LLAMA_CLOUD_API_KEY
file = client.files.create(file="./10-K-excerpt.pdf", purpose="parse")
result = client.parsing.parse(
file_id=file.id,
tier="agentic_plus", # fast | cost_effective | agentic | agentic_plus
version="latest",
expand=["markdown", "items"], # items carry bounding boxes per element
)
print(result.markdown.pages[0].markdown) # tables and charts as Markdown2. Claude Opus 5.5
ParseBench-2: 85.7 overall
Claude Opus 5.5 is the most accurate frontier model you can point at a document image today. Send it a page as an image and a prompt asking for Markdown, and it returns text at 92.7 and tables at 94.1 on ParseBench-2, within half a point of LlamaParse Agentic Plus on tables. Its visual grounding score of 86.2 is also the highest of any frontier model except Gemini 4, which matters if you ask it to return coordinates.
Two things hold it back as a parser. Charts score 74.1, well behind the 92.6 that Agentic Plus reaches, and the cost works out at $65 per 1,000 pages in the ParseBench-2 run, the highest of any ranked configuration. Claude Sonnet 5.5 at $2 per million input tokens drops to 73.8 overall, and Haiku 5.5 to 70.1, so the accuracy is tied to the flagship price.
What you get
- Image input through the Claude API, with a context window up to 1M tokens depending on model.
- Whatever structure you prompt for: Markdown, JSON, HTML tables. No bounding boxes or confidence unless you design the prompt to return them.
- ParseBench-2: 85.7 overall, 94.1 tables, 74.1 charts, 86.2 visual grounding.
Best for
- Low-volume, high-value pages where a reasoning model's judgement on a messy table is worth the price.
- Question answering over a document rather than converting it, where you need the model to read and reason in one call.
Watch out for
- Per-token pricing means per-page cost varies with page density and your prompt. Measure it on your own documents.
- Output is only as repeatable as your prompt. Two runs can format the same table differently.
Integration and cost
- Integration: Claude API and SDKs; available through Amazon Bedrock and Google Cloud Vertex AI.
- Runs in your network: no. Hosted API only, with cloud-marketplace options for data residency.
- Cost: $4 per million input tokens and $20 per million output tokens, 50% off with batch processing. ParseBench-2 measured $65 per 1,000 pages.
3. Gemini 4
ParseBench-2: 84.2 overall
Gemini 4, released on October 13, 2026 as Gemini 4 Argon, is the best value among frontier models for documents. It scores 84.2 overall on ParseBench-2, 1.5 points behind Claude Opus 5.5, and cost $27 per 1,000 pages in that run against Claude's $65. Its visual grounding score of 88.1 is the highest of any frontier model, so when you ask for coordinates it tends to get them right.
The weak dimension is charts, at 68.2, and the cheaper Gemini tiers fall off faster than the price does. Gemini 3 Flash scores 76.6 overall at about $9 per 1,000 pages, and the Flash-Lite models land between 61 and 63 with chart scores in the single digits. For text-heavy pages Flash is a bargain; for anything with a chart, you're paying for Gemini 4 or moving to a parser.
What you get
- Image, PDF and video input through the Gemini API, with a 1M-token combined input and output limit.
- Any text structure you prompt for; no element bounding boxes or confidence by default.
- ParseBench-2: 84.2 overall, 92.8 tables, 68.2 charts, 88.1 visual grounding.
Best for
- High-volume, text-heavy documents on Google Cloud where Flash tiers cover most pages.
- Long multi-page files where the 1M-token limit lets you send the whole document at once.
Watch out for
- Paid tier only: there's no free tier for Gemini 4.
- Chart accuracy drops sharply below the flagship, which is where most financial documents will bite.
Integration and cost
- Integration: Gemini API, Google AI Studio and Vertex AI.
- Runs in your network: no. Hosted API only.
- Cost: $4 per million input tokens and $20 per million output at Google's standard rate, cached input 95% off. ParseBench-2 measured $27 per 1,000 pages for Gemini 4 and $9 for Gemini 3 Flash.
4. GPT-5.6
ParseBench-2: 77.1 overall
OpenAI's models read documents competently and ground them poorly. GPT-5.6 scores 77.1 overall on ParseBench-2 with 87.9 on formatting and 76.3 on charts, ahead of both Claude Opus 5.5 and Gemini 4 on charts, but only 62.5 on visual grounding. If your pipeline needs to know where a value sits on the page, that's the number to look at. The newer GPT-6.1 Terra scored 74.3 and the small GPT-6 Luna 73.5, so the family is consistent rather than improving on documents.
Cost is the reason to consider it anyway. GPT-6 Luna parsed ParseBench-2 pages at about $1 per 1,000, the cheapest frontier configuration measured, and GPT-6.1 Terra at $12. For a first pass over clean text with a parser behind it for the hard pages, those are useful numbers.
What you get
- Image input through the OpenAI API across the GPT-5.6 and GPT-6 families, with a 1,050,000-token context window on GPT-5.6.
- Any text structure you prompt for; no element bounding boxes or confidence by default.
- ParseBench-2: GPT-5.6 77.1 overall, 79.6 tables, 76.3 charts, 62.5 visual grounding; GPT-6.1 Terra 74.3; GPT-6 Luna 73.5.
Best for
- Cheap first-pass conversion of clean, text-heavy pages at volume.
- Teams already on the OpenAI platform that want one vendor for reading and reasoning.
Watch out for
- Grounding scores in the 50s and 60s across the family. Don't rely on these models for coordinates.
- Tables at 79.6 trail Claude, Gemini and LlamaParse by 13 to 15 points.
Integration and cost
- Integration: OpenAI API and SDKs; also on Azure OpenAI.
- Runs in your network: no. Hosted API only.
- Cost: billed per token. GPT-6.1 Terra is $0.50 per million input tokens and $2.50 output, GPT-6 Luna $0.05 and $0.25, and GPT-5.6's current rate is on OpenAI's pricing page. In the ParseBench-2 run, GPT-5.6 cost $31 per 1,000 pages, GPT-6.1 Terra $12 and GPT-6 Luna $1.
5. MinerU
ParseBench-2: 75.6 overall
MinerU is OpenDataLab's open-source document parser, and MinerU2.5 is the 1.2B-parameter vision language model inside it. The model works in two stages: a layout pass on a downsampled page, then recognition on native-resolution crops of each region. That design is why it parses well for its size. MinerU2.5 Pro scored 75.6 overall on ParseBench-2 with 82.2 on visual grounding, the best grounding of any open-weight model after Infinity-Parser2.
What you're adopting is a toolkit as much as a model. The project, now at version 4.1, ships CLI and Python APIs, a self-hosted server, and renders to Markdown, HTML, LaTeX, DOCX and JSON. It also runs on CPU in its lighter tiers, with the VLM tier needing an 8 GB-plus NVIDIA GPU.
What you get
- A 1.2B document VLM with pipeline, vLLM, LMDeploy and llama.cpp backends; quality tiers from flash to advanced.
- Markdown, HTML, LaTeX, DOCX, EPUB, PDF and structured JSON output with layout boxes.
- ParseBench-2: 75.6 overall, 80.6 tables, 64.0 charts, 82.2 visual grounding.
Best for
- On-premises parsing of reports and papers where tables and formulas matter and the budget is GPU time.
- Teams that want a full self-hosted pipeline, not just weights.
Watch out for
- Some ParseBench-2 pages failed operationally and scored zero, so the 75.6 understates the model on pages it completes and overstates reliability.
- Charts at 64.0 and forms at 71.1 are the gaps; you'll want a review step on both.
Integration and cost
- Integration: Python package and CLI; self-hosted HTTP server with a tasks API; optional hosted parsing at mineru.net.
- Runs in your network: yes. Self-hosted on CPU or an 8 GB-plus NVIDIA GPU.
- Cost: free to run; you pay for hardware. ParseBench-2 didn't measure serving cost for open-weight models.
6. PaddleOCR-VL
ParseBench-2: 70.3 overall
PaddleOCR-VL is Baidu's 0.9B-parameter document model, built from a dynamic-resolution vision encoder and a 0.3B ERNIE language model. Version 1.6, released on May 28, 2026, is the current one. It's the smallest capable document VLM here and the most multilingual, with 109 languages supported since the first release and seal, stamp and text-spotting tasks added since.
Baidu reports 96.33% on OmniDocBench v1.6, a benchmark the model's team helps shape. On ParseBench-2, which it didn't train toward, PaddleOCR-VL 1.6 scored 70.3 overall with 73.0 on tables and 70.1 on visual grounding. Both numbers are true; the gap between them is the argument for testing on your own pages.
What you get
- A 0.9B Apache 2.0 model that handles text, tables, formulas, charts, seals and text spotting across 109 languages.
- Markdown and JSON with layout, served through PaddleOCR, a vLLM server, Transformers, or quantised builds for llama.cpp, Ollama and LM Studio.
- ParseBench-2: 70.3 overall, 73.0 tables, 57.5 charts, 70.1 visual grounding.
Best for
- Multilingual document volumes, especially Chinese, Japanese and other non-Latin scripts.
- Edge or single-GPU deployments where a 0.9B model is the whole budget.
Watch out for
- The 26-point difference between its OmniDocBench and ParseBench-2 scores. Pick the benchmark that looks like your documents.
- Full-page parsing needs the PaddleOCR pipeline or vLLM; the Transformers route covers element recognition only.
Integration and cost
- Integration: PaddleOCR CLI and Python API; vLLM server; Hugging Face Transformers; GGUF quantisations.
- Runs in your network: yes. Self-hosted, with CPU builds available.
- Cost: free, Apache 2.0; you pay for hardware.
7. Qwen3.8
ParseBench-2: 61.6 overall
Qwen3.8-27B is Alibaba's general-purpose open-weight vision language model, released in August 2026 under Apache 2.0. It reads images and video natively, carries a 262K-token context, and is the model most often named when developers ask for the best local VLM. On general vision tasks that reputation holds. On documents, ParseBench-2 scored it 61.6 overall in its default configuration, with text at 81.6 but tables at 36.7.
That table score is the story. A general VLM describes a table well and reconstructs it badly, and Qwen3.8 is the clearest example on this list. Alibaba's own card reports 91.1 on OmniDocBench 1.5, and a third-party ParseBench-2 run with thinking enabled reached 73.4, so configuration moves the number a lot. For documents, treat it as a reasoning layer on top of a parser's output rather than the parser.
What you get
- A 27B dense, natively multimodal model under Apache 2.0, with image and video input and a 262K context expandable to 1M.
- Any structure you prompt for; no element bounding boxes or confidence by default.
- ParseBench-2: 61.6 overall, 81.6 text, 36.7 tables, 63.5 visual grounding.
Best for
- Reasoning over parsed documents on your own hardware: summarising, classifying, answering questions.
- Mixed workloads where one local model has to handle images, video and text.
Watch out for
- Tables at 36.7 on ParseBench-2. Don't use it to reconstruct financial tables without a parser in front.
- A 27B dense model needs serious GPU memory; the quantised builds trade accuracy for fit.
Integration and cost
- Integration: Hugging Face Transformers, vLLM, SGLang, llama.cpp, Ollama and LM Studio builds; Qwen Cloud API announced as coming.
- Runs in your network: yes. Self-hosted.
- Cost: free, Apache 2.0; you pay for hardware.
8. olmOCR 2
ParseBench-2: Not ranked. ParseBench-2 evaluated olmOCR 2 but left it out of the ranked table because it returns no element boxes.
olmOCR 2 is the Allen Institute for AI's 7B-parameter document VLM, trained to turn PDF pages into clean, linearised Markdown. It's the model to look at when the goal is text for training or indexing rather than a faithful layout: equations, tables and handwriting come through, reading order is good, and the whole thing is Apache 2.0 with the training data and benchmark published alongside it.
Ai2's own olmOCR-Bench puts the October 2025 release at 82.4, and the project claims under $200 per million pages on your own GPUs. The trade-off is what it doesn't return. There are no bounding boxes, which is why ParseBench-2 couldn't rank it, and no confidence scores, so you can't trace a value back to the page or route uncertain ones to a person.
What you get
- A 7B Apache 2.0 VLM with pip, Docker and vLLM paths, and FP8 weights for faster serving.
- Markdown per page, plus Dolma-format output for training corpora.
- olmOCR-Bench 82.4; tables 84.9 and old-scan math 82.3 on the project's own benchmark.
Best for
- Converting large PDF collections to text for search, indexing or model training.
- Academic and scanned-book corpora with equations and older print.
Watch out for
- No element boxes and no confidence. If you need evidence per value, this isn't the model.
- Needs a recent NVIDIA GPU with 12 GB-plus of memory and about 30 GB of disk for the model image.
Integration and cost
- Integration: Python package, Docker image, vLLM server; hosted on several inference providers at around $0.07 to $0.10 per million input tokens.
- Runs in your network: yes. Self-hosted.
- Cost: free, Apache 2.0; Ai2 estimates under $200 per million pages on your own hardware.
9. DeepSeek-OCR 2
ParseBench-2: Not ranked.
DeepSeek-OCR 2, released in January 2026, is a 3B-parameter Apache 2.0 model built around an idea rather than a benchmark: compress a page into as few vision tokens as possible, then decode it. The default mode spends between 256 and about 1,100 visual tokens on a page depending on its size, which is a fraction of what a general VLM uses, and that's what makes it cheap to run at scale.
The first DeepSeek-OCR, from October 2025, is the one most people have tried; version 2 keeps the token budget and changes the visual encoder to read in a more human-like order. DeepSeek reports 76.3 on olmOCR-Bench for version 2. It hasn't published a ParseBench-2 result, and it returns Markdown or free text rather than element boxes, so layout fidelity is something you'll have to check yourself.
What you get
- A 3B Apache 2.0 model with vLLM and Transformers inference and a PDF concurrency mode.
- Markdown, plain text, figure parsing and tag-based grounding on request.
- A small vision-token budget per page, which is the whole efficiency argument.
Best for
- Bulk conversion where GPU hours are the constraint and pages are mostly text.
- Research into long-context compression, which is what the model was built to study.
Watch out for
- No ParseBench-2 result and no OmniDocBench figure on the model card; benchmark it before you commit.
- Setup is pinned to specific CUDA, torch and vLLM versions; expect some environment work.
Integration and cost
- Integration: Hugging Face weights, vLLM 0.8.5 scripts, Transformers.
- Runs in your network: yes. Self-hosted.
- Cost: free, Apache 2.0; you pay for hardware.
10. dots.mocr
ParseBench-2: Not ranked.
dots.mocr is rednote's 3B document VLM, the renamed successor to dots.ocr, which arrived in July 2025 as a 1.7B model that did layout detection and recognition in a single pass. The rename came with the March 2026 release, along with a companion model that outputs SVG for structured graphics. It's MIT licensed, though the repository also ships its own licence agreement, so read both.
What sets it apart is the output. Each page comes back as layout JSON with bounding boxes, categories and text, plus Markdown, which puts it closer to a parser than most open-weight models. rednote reports an Elo of 1059 on OmniDocBench v1.5. There's no ParseBench-2 result, and the team's own notes say complex tables and formulas are still difficult.
What you get
- A 3B model that returns layout JSON with boxes and categories, Markdown, and a layout visualisation per page.
- vLLM integration since 0.11.0, a Docker image, Transformers, and a separate SVG-output model.
- OmniDocBench v1.5 Elo 1059, self-reported.
Best for
- Self-hosted pipelines that need element boxes from a small model.
- Pages with structured graphics where the SVG companion model helps.
Watch out for
- Complex tables and formulas remain weak by the team's own account.
- Two licence files in one repository; have legal read the agreement before production use.
Integration and cost
- Integration: vLLM, Transformers, Docker; weights on Hugging Face and ModelScope.
- Runs in your network: yes. Self-hosted, with a CPU inference path.
- Cost: free, MIT; you pay for hardware.
Choosing a vision language model for document parsing
Is a VLM better than OCR?
For anything with a layout, yes. Classic OCR returns characters and coordinates and leaves tables, reading order and charts to you. A vision language model reads the page as an image and returns structure, which is why every ranked model here beats Azure Document Intelligence's layout OCR, at 66.1 on ParseBench-2, on tables. OCR still says where each word is and how sure it was. LlamaParse combines both: the VLM's reading with a bounding box for every element, which is how it reaches 87.5 on visual grounding.
Which open-source vision language models are best for documents?
On ParseBench-2, the ranked open-weight models are Infinity-Parser2 Pro at 77.1, LightOnOCR-3 4B at 77.0, Chandra 2 at 76.2, MinerU2.5 Pro at 75.6, Nemotron Parse 3.5 at 73.5, PaddleOCR-VL 1.6 at 70.3 and Qwen3.8 27B at 61.6. Models trained on documents beat general VLMs by 10 to 15 points, and all of them trail the hosted parsers and frontier APIs. If you must self-host, MinerU gives you the fullest pipeline, PaddleOCR-VL the widest language coverage and olmOCR 2 the cleanest text. Plan a review step for charts.
Should you call a frontier model directly or use a document parser?
Call the model directly when volume is low and the task is reading plus reasoning in one step. Use a parser when you need the same output every time, a bounding box per value, and a per-page price you can forecast. The numbers make the case: Claude Opus 5.5 reaches 85.7 overall at $65 per 1,000 pages on ParseBench-2; LlamaParse Agentic Plus reaches 89.8 at $28 and Cost Effective 84.6 at $4. The parser is more accurate and cheaper because it runs vision models only where a page needs them.
How do you benchmark a VLM on your own documents?
Pick 50 of your worst pages, not typical ones: dense tables, charts, scans, multi-column layouts. Run each finalist and score three things. Did the values you'd act on come back correct, especially numbers in tables and charts. Can you find each value on the page from the output, which a bounding box gives you and a bare Markdown string doesn't. And what did the run cost per page. ParseBench-2's 4.3k cases are public if you want a larger set, and the technical report explains how each dimension is scored.
Final thoughts
Every model here can read a clean page. The ones that earn a place in a production pipeline reconstruct tables and charts correctly, tell you where each value came from, and cost an amount you can forecast per page.
LlamaParse is where we'd start: the top ParseBench-2 score at 89.8, a bounding box on every element, four tiers from a no-model Fast pass to Agentic Plus, and 10,000 free credits a month to prove it on your files. Claude Opus 5.5 is the frontier model to call directly when a hard page is worth $0.065, Gemini 4 when you want frontier accuracy at $27 per 1,000, and MinerU or PaddleOCR-VL when the documents can't leave your network. Qwen3.8 belongs behind a parser, reasoning over what it returns.
Then test it the only way that counts: 50 of your worst pages through each finalist, scored on the values your application needs, with a note of how many came back with a location you could check.
Frequently asked questions
What is the best vision language model for document parsing?
For production document parsing, LlamaParse: it runs vision models and parsing agents over each page, returns Markdown or JSON with a bounding box on every element, holds the top ParseBench-2 score at 89.8, and includes 10,000 free credits a month. The exceptions depend on your constraints:
| Your situation | Start with | Why |
|---|---|---|
| Financial reports with charts and nested tables | LlamaParse Agentic Plus | Leads ParseBench-2 on formatting, tables and charts; 89.8 overall |
| High volume, mostly text, fixed budget | LlamaParse Cost Effective | 84.6 overall at $4 per 1,000 pages in the ParseBench-2 run |
| Values an agent will act on | LlamaParse | Bounding box and page for every element; typed fields with confidence |
| A few hard pages, read and reasoned in one call | Claude Opus 5.5 | Best frontier accuracy: 85.7 overall, 94.1 on tables |
| Frontier accuracy on Google Cloud at lower cost | Gemini 4 | 84.2 overall at $27 per 1,000 pages |
| Documents that can't leave your network | MinerU | Full self-hosted pipeline; 75.6 overall, 82.2 grounding |
| Non-Latin scripts on a single GPU | PaddleOCR-VL 1.6 | 0.9B model, 109 languages, Apache 2.0 |
| Clean text from large PDF collections for indexing | olmOCR 2 | Markdown at under $200 per million pages, Apache 2.0 |
What are the best document parsers built on vision language models?
LlamaParse leads ParseBench-2 among hosted parsers at 89.8 overall, with its Agentic and Cost Effective tiers at 86.0 and 84.6. Among open-source parsers built on VLMs, MinerU scores 75.6, PaddleOCR-VL 1.6 70.3, and olmOCR 2 and dots.mocr publish strong results on their own benchmarks without a ParseBench-2 ranking. Frontier models called directly sit between the two groups: Claude Opus 5.5 at 85.7, Gemini 4 at 84.2, GPT-5.6 at 77.1. Our glossary entry on VLM document parsing explains how these systems differ from OCR.
Is ChatGPT a vision language model?
Yes. The GPT models behind ChatGPT accept images alongside text, which is the definition of a vision language model, and the same is true of Claude and Gemini. Paste a page into any of them and you'll get a reading back. What you won't get is a bounding box per value, a confidence score, or a repeatable per-page cost, and on ParseBench-2 the best of them scores 85.7 against LlamaParse's 89.8. They're a fine way to prototype and a poor way to run a pipeline.
Which AI has the best vision model for documents?
Anthropic's Claude Opus 5.5 scores highest among frontier models on ParseBench-2 at 85.7, with Google's Gemini 4 close behind at 84.2 and cheaper. OpenAI's GPT-5.6 scores 77.1. All three trail LlamaParse Agentic Plus at 89.8, which uses vision models inside a parsing pipeline rather than as the whole answer. For general image understanding rather than documents, the public Vision Arena leaderboard is the better reference.
What is the best open-source vision language model for OCR?
It depends on what you mean by OCR. For clean Markdown from PDFs at scale, olmOCR 2, with an 82.4 on its own benchmark and an Apache 2.0 licence. For a full self-hosted parsing pipeline with layout boxes, MinerU at 75.6 on ParseBench-2. For 109 languages in a 0.9B model, PaddleOCR-VL 1.6. For a general local VLM that also reads documents, Qwen3.8-27B, with the caveat that its ParseBench-2 table score is 36.7. If your pages are photos rather than PDFs, see our image to text converter comparison.
How much does it cost to parse documents with a vision language model?
ParseBench-2 measured it per configuration. Frontier models called directly ran from about $1 per 1,000 pages for GPT-6 Luna to $65 for Claude Opus 5.5, with Gemini 4 at $27 and GPT-5.6 at $31. LlamaParse ran at $4 for Cost Effective, $13 for Agentic and $28 for Agentic Plus; on list pricing those tiers are $4.50, $15 and $37.50 per 1,000 pages, with 10,000 free credits a month. Open-weight models cost GPU time: olmOCR 2's team estimates under $200 per million pages.





