Best AI for extracting data from legal documents
Legal Benchmarks: Data Extraction
| Option and model | Task pass rate | Cost per benchmark task |
|---|---|---|
| Highest scoreClaude Opus 5.5 Anthropic | 76.7% | $2.71 |
| Lower-cost optionClaude Fable 5.1 Anthropic | 63.3% | $1.15 |
Source: Legal Benchmarks ↗ · as of Nov 2026 · last checked 17 Nov 2026
Scoring rule: Highest score is the top task pass rate on this leaderboard. Lower-cost option is the best-scoring model whose published cost is below the top scorer’s; equal scores go to the lower cost. Cost is API cost per benchmark task as published by the source, not a subscription price.
Claude Opus 5.5 (Anthropic) has the top score on Legal Benchmarks: Data Extraction: 76.7% (task pass rate). It costs $2.71 API cost per benchmark task; the best-scoring cheaper option is Claude Fable 5.1 (Anthropic) at 63.3% for $1.15 per task. What the test covered: Data extraction tasks written by lawyers, each run twice; a task passes only if every criterion is met both times.
- Legal Benchmarks: Data Extraction: published by Legal Benchmarks ↗ · benchmark dated Nov 2026 · last checked 17 Nov 2026
Citation
Legal Benchmarks: Data Extraction, task pass rate: Claude Opus 5.5 (Anthropic) had the top reported score at 76.7%. Original source: Legal Benchmarks, https://www.legalbenchmarks.ai/leaderboard/models, leaderboard dated Nov 2026. Via ChatJudge: https://chatjudge.ai/best-ai/data-extraction-from-documents, checked 17 Nov 2026.
Test the three leading models on your own documents. Compare models on my documents · US$299 How it works
ChatJudge prompt
Try it: Extract parties, dates and obligations, with the source passage for each
A ChatJudge prompt, written by ChatJudge for this guide (not credited to anyone else).
Extract the key terms from the legal document below. Return a table with one row per item and these columns: Field, Value, Source passage, Location. Extract: 1. Parties: each party's full legal name and its role (for example buyer, supplier, licensor). 2. Dates: the effective date, term and end date, renewal and notice deadlines, payment dates and any other deadline. 3. Obligations: each obligation, who owes it, to whom, and by when. Rules: - For every value, quote the exact source passage word for word in quotation marks, and give its clause, section or page number. - If something is not in the document, write "Not stated". Never infer, estimate or fill a gap. - End with a short "Check with a lawyer" list of anything ambiguous or contradictory. Document: [PASTE THE CONTRACT OR LEGAL DOCUMENT]
Paste it into two of the leaders above, run it on the same document, and compare. Check every answer against the source before you rely on it.
Benchmark data
Results from published AI leaderboards. Compare models within each benchmark; scores from different tests aren’t directly comparable. How we collect these results
Legal Benchmarks: Data Extraction
Data extraction tasks written by lawyers, each run twice; a task passes only if every criterion is met both times. Run by Legal Benchmarks. Leaderboard dated Nov 2026. Last checked 17 Nov 2026.
- Anthropic
- OpenAI
- xAI
- Meta
- DeepSeek
- Alibaba
- Z.ai
- Moonshot
- MiniMax
- Mistral
Even the top entry passed 76.7% of tasks under this rule, so it missed at least one requirement on the rest.
All scores
| Rank | Model | Provider | Task pass rate | Criteria pass rate | Form | Cost per task (USD) |
|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | Anthropic | 76.7% | 91.1% | 82.2% | $2.71 |
| 2 | Claude Fable 5.2 | Anthropic | 66.7% | 89.2% | 79.3% | $5.90 |
| 3 | Claude Fable 5.1 | Anthropic | 63.3% | 89.9% | 70.6% | $1.15 |
| 4 | Claude Opus 5 | Anthropic | 56.7% | 89.2% | 67.2% | $0.75 |
| 4 | GPT-6 Astra | OpenAI | 56.7% | 88.4% | 83.5% | $0.93 |
| 4 | Qwen 3.8 Flash | Alibaba | 56.7% | 85.7% | 64.4% | $0.030 |
| 4 | Claude Haiku 5.5 | Anthropic | 56.7% | 82.3% | 77.4% | $0.20 |
| 4 | Qwen 3.8 Max | Alibaba | 56.7% | 77.6% | 62.2% | $1.51 |
| 4 | GLM-5.3 | Z.ai | 56.7% | 72.4% | 67.2% | $0.12 |
| 10 | Muse Spark 1.3 | Meta | 53.3% | 86.5% | 67.4% | $0.29 |
Show 17 more rows
| Rank | Model | Provider | Task pass rate | Criteria pass rate | Form | Cost per task (USD) |
|---|---|---|---|---|---|---|
| 10 | Gemini 4 Argon | 53.3% | 86.2% | 66.5% | $0.27 | |
| 10 | GPT-6.1 Sol | OpenAI | 53.3% | 84.5% | 83.0% | $0.13 |
| 10 | Grok 4.6 | xAI | 53.3% | 78.3% | 69.3% | $0.45 |
| 10 | Gemini 3.8 Flash | 53.3% | 71.7% | 66.7% | $0.21 | |
| 15 | MiniMax M3 | MiniMax | 50.0% | 79.8% | 72.8% | $0.030 |
| 15 | DeepSeek V4 Pro | DeepSeek | 50.0% | 57.1% | 67.4% | $0.15 |
| 17 | GPT-5.6 Sol | OpenAI | 46.7% | 82.0% | 78.9% | $0.20 |
| 17 | Claude Sonnet 5 | Anthropic | 46.7% | 78.8% | 70.6% | $0.26 |
| 19 | Kimi K3 | Moonshot | 43.3% | 81.8% | 75.0% | $0.49 |
| 20 | Claude Sonnet 5.5 | Anthropic | 40.0% | 83.3% | 77.6% | $0.92 |
| 20 | GPT-5.6 Terra | OpenAI | 40.0% | 78.1% | 80.9% | $0.12 |
| 20 | GPT-5.6 Luna | OpenAI | 40.0% | 77.3% | 79.6% | $0.020 |
| 20 | GPT-6.1 Terra | OpenAI | 40.0% | 73.2% | 75.4% | $0.070 |
| 20 | Mistral Large 4 | Mistral | 40.0% | 65.8% | 60.9% | $0.13 |
| 25 | DeepSeek V4 Flash | DeepSeek | 36.7% | 36.9% | 43.3% | $0.020 |
| 26 | GLM-5.3 Flash | Z.ai | 30.0% | 46.6% | 47.2% | $0.030 |
| 27 | Mistral Medium 3.5 | Mistral | 20.0% | 43.1% | 62.2% | $0.51 |
Full Legal Benchmarks: Data Extraction leaderboard: every model, all metrics and the table →
What this benchmark tells you
- How often a model completes lawyer-written data extraction tasks on native source documents with every applicable criterion met, in both of two attempts.
- How many individual criteria it meets, its form score and its cost per task, as published.
What it doesn’t tell you
- The tasks are legal documents, so results on invoices, receipts, forms or interview notes may differ.
- It does not tell you how a model handles your file formats or scan quality; test on a sample of your own documents.
Private model comparison
Which model extracts the fields your team needs, and what does it cost per document?
Both options compare the same three named models, Anthropic’s Claude Opus 5.5 (claude-opus-5-5), OpenAI’s GPT-6 Astra (gpt-6-astra), Google’s Gemini 4 Carbon (gemini-4-carbon), on the same five redacted documents of yours, and include a private report covering essential-field accuracy, representative errors and cost per document.
US$299 one time
Model Comparison
Find the best model for your documents.
- Pick an invoice, purchase-order or contract template, adjust the fields and mark the ones that must be right.
- Confirm the expected values: highlight them in the text, import a spreadsheet, or mark a field absent.
- An automated report shows which models met your requirements, including ties and cases where none did, with cost per document.
- You see the delivery deadline before you pay. If we miss it, the refund is automatic.
Compare models on my documents · US$299 Try it with synthetic examples
Online checkout is not open yet. You can do the full setup now and send it as a request: nothing is charged.
Claude Opus 5.5, GPT-6 Astra and Gemini 4 Carbon all captured every field you marked essential on these five documents, so they tie on your requirements.
Of those, Gemini 4 Carbon costs least, at $0.0081 per document against $0.0184 for Claude Opus 5.5 and $0.0239 for GPT-6 Astra.
| Essential field | Claude Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|
| Invoice number | 5/5 | 5/5 | 5/5 |
| Invoice date | 5/5 | 5/5 | 5/5 |
| Supplier name | 5/5 | 5/5 | 5/5 |
| Purchase-order number | 5/5 | 5/5 | 5/5 |
| Total amount due | 5/5 | 5/5 | 5/5 |
| All essential values | 25/25 · met | 25/25 · met | 25/25 · met |
US$499 one time
Reviewed Comparison
Find out why models miss key fields and test a fix.
Everything in the Model Comparison, plus a human reviewer who checks the essential-field results against the source text, investigates the most consequential error or ambiguity, tests one revised extraction instruction and reruns the same three models on the same five documents. You get the tested instructions, a before-and-after scorecard showing any regressions, a one-page recommendation you can share, and an optional 20-minute walkthrough. The answer key and scoring rules stay fixed, and we report the actual result even if the revision does not help.
No charge for a request. Before any payment we confirm the reviewer, the deliverables and a delivery date in writing. Already bought the US$299 comparison? You can add a reviewer from your results page and pay only the difference.
Who sees your files and when they are deleted: privacy and deletion · privacy policy · terms. Prefer to test it yourself? The free prompt above is the same kind of instruction.
More prompts for extracting data from legal documents
Run the same prompt on your own material with two models and compare the results.
Questions people ask
Which AI has the highest reported score for extracting data from legal documents?
As of Nov 2026, the highest reported score on Legal Benchmarks: Data Extraction is Claude Opus 5.5 (Anthropic) at 76.7%, per Legal Benchmarks.
How recent are these results?
Legal Benchmarks: Data Extraction: leaderboard dated Nov 2026; we last checked it on 17 Nov 2026.
What does the top-scoring model cost to run on these tests?
On Legal Benchmarks: Data Extraction, Claude Opus 5.5 costs $2.71 (cost per task), as published by Legal Benchmarks. That is API cost on the benchmark’s own tasks, not a subscription price.
Who runs Legal Benchmarks: Data Extraction?
Legal Benchmarks: Data Extraction is run by Legal Benchmarks. ChatJudge does not run the tests; we copy the published scores with a link to the source.
What don't these benchmarks tell you about extracting data from legal documents?
The tasks are legal documents, so results on invoices, receipts, forms or interview notes may differ. It does not tell you how a model handles your file formats or scan quality; test on a sample of your own documents.
