Skip to main content
BenchLM
Data
Specialist vertical · BenchAlign v6.0 · 87 ranked (49 Supported, 38 Estimated)
Scored from the dataset of

Documents and extraction leaderboard

As of November 21, 2026, GPT-6.2 Astra leads our documents and extraction leaderboard at 86.5 (Estimated, $8/$40 per 1M tokens), ahead of Claude Fable 5.2 at 81.8. Gemini 4 Carbon is the highest Supported row at #3 (80.7); Qwen3.8 Max leads open-weight models at #19 (66.2). We rank 87 models (49 Supported, 38 Estimated) across 14 domain protocols, and this vertical carries zero weight in the overall score.

Shortlist by workload and budget

Each pick comes directly from the live documents & extraction rows below, with listed API rates and the case where another option wins.

Highest point score#1 · Estimated

OpenAI · $8/$40 per 1M tokens

Carries an Estimated label; pick Gemini 4 Carbon (#3, 80.7) when you require multi-source Supported coverage.

Highest Supported row#3 · Supported

Google · $2/$10 per 1M tokens

Trails GPT-6.2 Astra by 5.8 points, so you trade peak point score for broader independent verification.

Best API value#3 · Supported

Google · $2/$10 per 1M tokens

Keeps 93% of the top score (80.7 vs 86.5), but gives up 5.8 points on the hardest tail cases.

Top open-weight model#19 · Estimated

Alibaba · Open weights · self-hostable

Trails GPT-6.2 Astra by 20.3 points (#19 overall here), so pick it when weight control or on-prem residency outweighs peak score.

Ranked models

Ordered by BenchAlign v6.0 point estimate as of November 21, 2026. An Estimated row backed by fewer than two source families never ranks above a row scoring within 2 points below it. Conditional ranges do not establish rank confidence.

Documents & Extraction LLM Leaderboard
RankModelScoreEvidence
#1OpenAI86.5Estimated
#2Anthropic81.8Estimated
#3Google80.7Supported
#4OpenAI80.0Supported
#5Google81.7Estimated
#6Google76.6Supported
#7Anthropic75.6Supported
#8OpenAI73.8Supported
#9Google75.7Estimated
#10Anthropic72.0Supported
#11OpenAI71.1Supported
#12OpenAI70.8Supported
#13OpenAI68.7Supported
#14Anthropic67.7Supported
#15Tencent67.6Supported
#16Anthropic66.9Supported
#17Google66.8Supported
#18Google66.6Supported
#19
Qwen3.8 MaxOpen weights
Alibaba
66.2Estimated
#20OpenAI65.9Supported
#21OpenAI65.6Supported
#22Anthropic65.1Supported
#23OpenAI64.7Estimated
#24Google64.5Supported
#25Google64.2Supported
#26OpenAI64.0Supported
#27
Kimi K3Open weights
Moonshot AI
63.2Supported
#28OpenAI63.0Supported
#29xAI62.7Supported
#30xAI62.5Supported
#31
Qwen3.8-Flash-NextOpen weights
Alibaba
62.4Estimated
#32Meta64.2Estimated
#33DeepSeek62.1Supported
#34OpenAI61.1Supported
#35
Qwen3.8-27BOpen weights
Alibaba
61.1Supported
#36Google60.3Supported
#37Anthropic59.9Supported
#38Anthropic59.6Supported
#39Anthropic59.5Supported
#40
Gemma 4 31BOpen weights
Google
59.4Supported
#41Anthropic59.1Supported
#42Meta61.0Estimated
#43OpenAI58.6Supported
#44
GLM-5.3-FlashOpen weights
Z.AI
58.6Supported
#45Google60.5Estimated
#46OpenAI60.3Estimated
#47Google56.0Supported
#48Anthropic55.7Estimated
#49Anthropic55.7Estimated
#50
Gemma 4 26B A4BOpen weights
Google
54.5Supported
#51Google53.4Supported
#52Anthropic52.7Supported
#53Anthropic52.7Supported
#54
GLM-5.3Open weights
Z.AI
53.9Estimated
#55Anthropic51.7Supported
#56Anthropic52.8Estimated
#57Anthropic52.8Estimated
#58Google49.9Supported
#59
Kimi K2.6Open weights
Moonshot AI
48.6Estimated
#60
DeepSeek-R1Open weights
DeepSeek
50.4Estimated
#61xAI48.1Estimated
#62OpenAI47.9Estimated
#63OpenAI47.7Supported
#64OpenAI48.8Estimated
#65Google46.6Supported
#66
Gemma 3 27BOpen weights
Google
48.5Estimated
#67xAI47.8Estimated
#68Anthropic44.6Supported
#69Anthropic44.6Supported
#70Mistral45.9Estimated
#71xAI44.8Estimated
#72xAI44.1Estimated
#73OpenAI43.7Estimated
#74xAI43.0Estimated
#75NVIDIA40.8Estimated
#76OpenAI39.2Supported
#77
Qwen3.5-35B-A3BOpen weights
Alibaba
37.7Estimated
#78OpenAI39.5Estimated
#79
DeepSeek V3.2Open weights
DeepSeek
39.5Estimated
#80
DeepSeek V3.1Open weights
DeepSeek
39.2Estimated
#81
Qwen3.6-35B-A3BOpen weights
Alibaba
36.8Supported
#82
Gemma 4 E2BOpen weights
Google
32.1Estimated
#83
DeepSeek V3Open weights
DeepSeek
31.8Estimated
#84
Gemma 4 E4BOpen weights
Google
31.2Supported
#85
Gemma 4 12BOpen weights
Google
27.7Estimated
#86
GPT-OSS 120BOpen weights
OpenAI
15.2Estimated
#87
GPT-OSS 20BOpen weights
OpenAI
11.9Estimated

Scored domain benchmarks (14)

Each percentage is the protocol's relative reference weight inside this vertical's calibrated BenchAlign v6.0 model, not a raw score average. Independent evaluator shares are capped at 40% whenever three or more independent evaluators contribute.

FACTS Grounding v2 checks whether long-form answers remain supported by supplied documents. Its revised judge ensemble and public/private results remain separate from version 1.

12% ref. weight

Scores document parsers and vision-language models on 4,317 pages from 2,645 documents in six dimensions: text, formatting, tables, charts, visual grounding and forms, with severity-aware scoring and the listed cost per page. A separate release from ParseBench-1; the scores are not comparable.

12% ref. weight

The first ParseBench release: scores document parsers and vision-language models on 2,078 human-verified enterprise pages across tables, charts, content faithfulness, semantic formatting and visual grounding, with the listed cost per page. ParseBench-2 is a separate benchmark with different scoring.

12% ref. weight

An agentic spatial reasoning benchmark reported as a normalized score.

10% ref. weight

A long-context benchmark that measures whether models can actually use extended context windows for reasoning and retrieval.

8% ref. weight

A benchmark for grounded reasoning over office-style documents, spreadsheets, charts, and business artifacts.

7% ref. weight

A scientific chart reasoning benchmark that tests whether models can understand, interpret, and reason about complex scientific visualizations including plots, diagrams, and data charts.

7% ref. weight

Tool-free variant of CharXiv that isolates raw visual reasoning ability without code execution or tool augmentation.

7% ref. weight

Scores schema-guided extraction from 370 enterprise documents and 4,869 pages on value F1, record completeness, word and page grounding and cost, split by short, medium and long documents.

7% ref. weight

A harder companion to ExtractBench: 100 synthetic documents and 2,622 pages across 86 document types that test document conventions, house vocabulary, counting rules, drawn relations and charts, scored by perfect-document rate.

6% ref. weight

A long-context benchmark for memory, retrieval, and multi-round coherence over large contexts.

4% ref. weight

A document understanding benchmark used in frontier-model comparison tables to measure extraction and grounded reasoning quality on complex documents.

4% ref. weight

A long-document multimodal benchmark for grounded reasoning over extended document contexts.

3% ref. weight

An OCR-focused benchmark for reading and extracting text from visually complex documents and images.

2% ref. weight

Structured extraction and OCR layout parsing separate models that look tied on chat

Fourteen document protocols feed this board, anchored by FACTS Grounding v2 (12%), ParseBench-2 (12%), ParseBench-1 (12%) alongside Blueprint-Bench 2, LongBench v2, OfficeQA Pro, ExtractBench, CharXiv, and OmniDocBench 1.5. Together they test three distinct failure modes: reading dense multi-column PDFs and engineering blueprints without losing coordinates, extracting nested JSON records that match a strict schema, and answering grounded questions without inventing numbers outside the source file.

General multimodal boards mix natural photos and UI screenshots with office files, which can hide weak table-cell alignment or silent field drops. Splitting document extraction into its own BenchAlign v6.0 vertical surfaces which models keep column headers, footnotes, and units intact across hundred-page packets.

Your OCR pipeline and page image budget change results as much as the model

Benchmark harnesses feed pre-sliced pages or standardized DPI renders. In production, skewed scans, handwritten margin edits, password-wrapped attachments, and 300-page credit agreements often break the ingestion step before the language model sees a token. A two-point lead here will disappear if your PDF splitter severs a table header from its rows.

Run a 50-document sample from your own archive, including your messiest scanned invoices or regulatory filings, and grade exact field match plus abstention on blank fields before committing to an annual contract.

Balancing extraction accuracy against per-page token cost

GPT-6.2 Astra (86.5, Estimated, $8/$40 per 1M tokens) edges out Claude Fable 5.2 (81.8, Supported, $8/$40 per 1M tokens) by just 4.7 points while costing four times as much per token. For high-volume invoice or filing ingestion, Claude Fable 5.2 (or Gemini 4 Carbon at #3, 80.7, $2/$10 per 1M tokens) usually delivers better unit economics unless a single missed clause triggers manual rework.

When customer contracts must stay on-premises, Qwen3.8 Max (#19, 66.2) is the strongest self-hostable option, though its 20.3-point gap behind the frontier argues for JSON-schema validation and confidence-based human review.

Questions

Which LLM is best for document extraction right now?

GPT-6.2 Astra holds #1 on our document extraction leaderboard at 86.5 (Estimated) as of November 21, 2026, ahead of Claude Fable 5.2 at 81.8 and Gemini 4 Carbon at 80.7. Gemini 4 Carbon (#3, 80.7) is the highest Supported option when independent multi-evaluator depth matters most.

What is the best open-weight LLM for document extraction?

Qwen3.8 Max from Alibaba leads open-weight models on this document extraction board at #19 with a calibrated score of 66.2 (Estimated). That trails GPT-6.2 Astra by 20.3 points, so choose Qwen3.8 Max when self-hosting, VPC data residency, or weight customization matters more than peak frontier accuracy.

Which API model offers the best value for document extraction?

Gemini 4 Carbon (#3, 80.7) is our top API value pick for document extraction, priced at $2/$10 per 1M tokens compared with GPT-6.2 Astra at $8/$40 per 1M tokens. You give up 5.8 calibrated points, making Gemini 4 Carbon the better buy whenever those extra points do not cut downstream retries.

Which benchmarks make up the document extraction leaderboard score?

BenchAlign v6.0 calibrates 14 open and primary domain protocols for this board, led by FACTS Grounding v2 (12%), ParseBench-2 (12%), ParseBench-1 (12%), Blueprint-Bench 2 (10%). Each percentage reflects share of the vertical's reference weight inside our item-response model rather than a raw score average, and no single independent evaluator can exceed 40% when three or more evaluators contribute.

How does the documents leaderboard differ from the multimodal leaderboard?

Our multimodal leaderboard blends general image understanding, visual puzzles, and UI screenshots with document tasks. This specialist board isolates PDF parsing, OCR fidelity, blueprint reading, long-document QA, and strict JSON schema extraction across 14 protocols, showing which models preserve tables and field types without visual-chat noise.

Do document extraction scores affect the overall ranking?

Document vertical scores stay separate and carry zero direct weight in the overall leaderboard, though shared grounded benchmarks still inform multimodal capability where applicable. Across 87 ranked models (49 Supported, 38 Estimated), test your own scanned PDFs and OCR settings before assuming benchmark parity.

Documents & extraction leaderboard updates

Weekly updates when new models or domain evaluations move the documents & extraction rankings.

Read a sample issue

Join 7,000+ subscribers.

One email each week. Unsubscribe anytime.