# Legal leaderboard

As of November 21, 2026, Claude Fable 5.2 leads our legal leaderboard at 86.1 (Estimated, $8/$40 per 1M tokens), ahead of Claude Opus 5.5 at 85.9. All 40 rows currently carry Estimated labels; Claude Opus 5.5 (#2, 85.9) is the lowest-cost model near the top at $4/$20 per 1M tokens; DeepSeek V4 Pro leads open-weight models at #13 (67.2). We rank 40 models across 11 domain protocols, and this vertical carries zero weight in the overall score.

## Shortlist by workload and budget

- **Highest point score:** [Claude Fable 5.2](/models/claude-fable-5-2) (#1, 86.1, Estimated, $8/$40 per 1M tokens) — Carries an Estimated label because independent multi-evaluator coverage on legal is still thin.
- **Closest contender:** [Claude Opus 5.5](/models/claude-opus-5-5) (#2, 85.9, Estimated, $4/$20 per 1M tokens) — Sits 0.2 points behind Claude Fable 5.2; test both on your own prompts when provider lock-in or latency matters.
- **Best API value:** [Claude Opus 5.5](/models/claude-opus-5-5) (#2, 85.9, Estimated, $4/$20 per 1M tokens) — Keeps 100% of the top score (85.9 vs 86.1), but gives up 0.2 points on the hardest tail cases.
- **Top open-weight model:** [DeepSeek V4 Pro](/models/deepseek-v4-pro) (#13, 67.2, Estimated, Open weights · $1.32/$3.96 per 1M tokens) — Trails Claude Fable 5.2 by 18.9 points (#13 overall here), so pick it when weight control or on-prem residency outweighs peak score.

## Ranked models

| Rank | Model | Creator | Score | Conditional range | Evidence | API ($/1M in/out) | Context | Source families |
|---:|---|---|---:|---:|---|---:|---:|---:|
| #1 | [Claude Fable 5.2](/models/claude-fable-5-2) | Anthropic | 86.1 | 74.8-97.5 | Estimated | $8.00 / $40.00 | 1M | 4 |
| #2 | [Claude Opus 5.5](/models/claude-opus-5-5) | Anthropic | 85.9 | 74.3-97.5 | Estimated | $4.00 / $20.00 | 1M | 4 |
| #3 | [Claude Fable 5.1](/models/claude-fable-5-1) | Anthropic | 81.1 | 70.8-91.3 | Estimated | $10.00 / $50.00 | 1M | 5 |
| #4 | [Claude Opus 5](/models/claude-opus-5) | Anthropic | 80.8 | 70.7-90.9 | Estimated | $5.00 / $25.00 | - | 5 |
| #5 | [GPT-6 Astra](/models/gpt-6-astra) | OpenAI | 72.7 | 62.4-82.9 | Estimated | $10.00 / $50.00 | 1.05M | 5 |
| #6 | [GPT-6.1 Sol](/models/gpt-6-1-sol) | OpenAI | 71.9 | 61.0-82.7 | Estimated | $2.00 / $10.00 | 1.05M | 4 |
| #7 | [Gemini 4 Argon](/models/gemini-4-argon) | Google | 71.2 | 60.3-82.2 | Estimated | $2.00 / $10.00 | 1M | 4 |
| #8 | [Claude Sonnet 5.5](/models/claude-sonnet-5-5) | Anthropic | 70.6 | 59.2-82.0 | Estimated | $2.00 / $10.00 | 1M | 4 |
| #9 | [Claude Haiku 5.5](/models/claude-haiku-5-5) | Anthropic | 70.3 | 59.4-81.3 | Estimated | $0.10 / $0.50 | 1M | 4 |
| #10 | [Grok 4.6](/models/grok-4-6) | xAI | 70.1 | 60.0-80.3 | Estimated | $2.00 / $6.00 | 500K | 5 |
| #11 | [Claude Fable 5](/models/claude-fable) | Anthropic | 71.6 | 61.1-82.0 | Estimated | $10.00 / $50.00 | 1M+ | 1 |
| #12 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | OpenAI | 68.2 | 58.0-78.3 | Estimated | $4.00 / $20.00 | 1.05M | 5 |
| #13 | [DeepSeek V4 Pro](/models/deepseek-v4-pro) | DeepSeek | 67.2 | 56.9-77.4 | Estimated | $1.32 / $3.96 | 1M | 5 |
| #14 | [Grok 4.5](/models/grok-4-5) | xAI | 69.0 | 58.5-79.5 | Estimated | $2.00 / $6.00 | 500K | 1 |
| #15 | [Gemini 3.8 Flash](/models/gemini-3-8-flash) | Google | 66.1 | 55.9-76.3 | Estimated | $0.75 / $3.75 | 1M | 5 |
| #16 | [Claude Sonnet 5](/models/claude-sonnet-5) | Anthropic | 64.7 | 54.7-74.7 | Estimated | $2.00 / $10.00 | 1M | 5 |
| #17 | [Claude Opus 4.8](/models/claude-opus-4-8) | Anthropic | 66.0 | 55.7-76.4 | Estimated | $5.00 / $25.00 | 1M | 1 |
| #18 | [GLM-5.3](/models/glm-5-3) | Z.AI | 63.0 | 51.6-74.3 | Estimated | Free | 1M | 4 |
| #19 | [MiniMax M3](/models/minimax-m3) | MiniMax | 62.4 | 51.9-73.0 | Estimated | $0.30 / $1.20 | 1M | 5 |
| #20 | [GPT-6.1 Terra](/models/gpt-6-1-terra) | OpenAI | 60.8 | 49.7-71.9 | Estimated | $1.00 / $5.00 | 1.05M | 4 |
| #21 | [GPT-5.6 Terra](/models/gpt-5-6-terra) | OpenAI | 60.5 | 50.4-70.7 | Estimated | $2.00 / $12.00 | 1.05M | 5 |
| #22 | [Kimi K3](/models/kimi-k3) | Moonshot AI | 60.5 | 49.9-71.0 | Estimated | $3.00 / $15.00 | 1.05M | 5 |
| #23 | [GPT-5.6 Luna](/models/gpt-5-6-luna) | OpenAI | 60.3 | 50.1-70.4 | Estimated | $0.20 / $1.20 | 1.05M | 5 |
| #24 | [Gemini 3.6 Flash](/models/gemini-3-6-flash) | Google | 62.1 | 51.6-72.5 | Estimated | $0.75 / $3.75 | 1M | 1 |
| #25 | [GLM-5.2](/models/glm-5-2) | Z.AI | 57.6 | 47.2-68.0 | Estimated | $1.40 / $4.40 | 1M | 1 |
| #26 | [Mistral Large 4](/models/mistral-large-4) | Mistral | 55.5 | 43.2-67.8 | Estimated | $0.70 / $2.10 | 1M | 4 |
| #27 | [Gemini 3.5 Flash](/models/gemini-3-5-flash) | Google | 51.8 | 41.4-62.2 | Estimated | $1.50 / $9.00 | 1M | 1 |
| #28 | [Gemini 3.1 Pro](/models/gemini-3-1-pro) | Google | 49.7 | 39.2-60.2 | Estimated | $2.00 / $12.00 | 1M | 1 |
| #29 | [Gemini 1.5 Pro](/models/gemini-1-5-pro) | Google | 47.5 | 29.0-66.0 | Estimated | $1.25 / $5.00 | 2M | 6 |
| #30 | [DeepSeek V3](/models/deepseek-v3) | DeepSeek | 45.3 | 27.7-62.9 | Estimated | $0.27 / $1.10 | 128K | 6 |
| #31 | [Mixtral 8x22B Instruct v0.1](/models/mixtral-8x22b-instruct-v0-1) | Mistral | 44.5 | 12.6-76.4 | Estimated | Free | 64K | 6 |
| #32 | [Claude 3.5 Sonnet](/models/claude-3-5-sonnet) | Anthropic | 44.4 | 27.5-61.3 | Estimated | $3.00 / $15.00 | 200K | 6 |
| #33 | [GPT-4o](/models/gpt-4o) | OpenAI | 43.4 | 26.1-60.8 | Estimated | $2.50 / $10.00 | 128K | 6 |
| #34 | [GPT-4 Turbo](/models/gpt-4-turbo) | OpenAI | 43.4 | 22.4-64.5 | Estimated | $10.00 / $30.00 | 128K | 6 |
| #35 | [Claude 3 Opus](/models/claude-3-opus) | Anthropic | 38.1 | 17.3-58.9 | Estimated | $15.00 / $75.00 | 200K | 6 |
| #36 | [GPT-4o mini](/models/gpt-4o-mini) | OpenAI | 37.5 | 20.8-54.2 | Estimated | $0.15 / $0.60 | 128K | 6 |
| #37 | [GLM-5.3-Flash](/models/glm-5-3-flash) | Z.AI | 37.3 | 25.1-49.4 | Estimated | Free | 1M | 4 |
| #38 | [Mistral Large 2](/models/mistral-large-2) | Mistral | 32.4 | 11.4-53.3 | Estimated | - | 128K | 6 |
| #39 | [Mistral Large 3](/models/mistral-large-3) | Mistral | 32.6 | 13.3-51.9 | Estimated | $0.50 / $1.50 | 128K | 1 |
| #40 | [Claude 3 Haiku](/models/claude-3-haiku) | Anthropic | 28.9 | 4.9-52.9 | Estimated | $0.25 / $1.25 | 200K | 6 |

## Scored domain benchmarks (11)

- Legal Benchmarks: Contract Workflows (criteria) · Legal Benchmarks (16% ref. weight): Rubric-level accuracy on commercial contract redlining, issue spotting, and clause revision.
- Legal Benchmarks: Contract Workflows (pass rate) · Legal Benchmarks (16% ref. weight): Strict all-criteria pass rate on commercial contract review and drafting workflows.
- DELTA Legal Research (criteria) · Legal Benchmarks (16% ref. weight): Rubric-level accuracy on statutory, regulatory, and case-law research memos.
- Legal Benchmarks: Data Extraction (criteria) · Legal Benchmarks (16% ref. weight): Field-level accuracy extracting obligations, dates, caps, and parties from agreements.
- Legal Benchmarks: Data Extraction (pass rate) · Legal Benchmarks (16% ref. weight): Strict document-level pass rate for structured legal clause and metadata extraction.
- HELM LegalBench: Abercrombie · Stanford CRFM HELM LegalBench (4% ref. weight): Trademark distinctiveness classification along the Abercrombie spectrum.
- HELM LegalBench: Corporate Lobbying · Stanford CRFM HELM LegalBench (4% ref. weight): Matching proposed federal bills to corporate lobbying disclosures and business relevance.
- HELM LegalBench: Decision Section Function · Stanford CRFM HELM LegalBench (4% ref. weight): Classifying rhetorical and procedural roles of judicial opinion paragraphs.
- HELM LegalBench: International Citizenship · Stanford CRFM HELM LegalBench (4% ref. weight): Statutory reasoning over nationality and citizenship rules across jurisdictions.
- Stanford HELM LegalBench (overall) · Stanford CRFM HELM LegalBench (4% ref. weight): Aggregate accuracy across the collaborative LegalBench legal reasoning suite on HELM.
- HELM LegalBench: PROA · Stanford CRFM HELM LegalBench (1% ref. weight): Identifying whether a statute creates an explicit private right of action.

## Commercial contract redlining, clause extraction, and statutory research

Eleven legal protocols set these rankings, led by Legal Benchmarks: Contract Workflows (criteria) (16%), Legal Benchmarks: Contract Workflows (pass rate) (16%), DELTA Legal Research (criteria) (16%) together with Legal Benchmarks extraction criteria and pass rates (80% combined reference share across transactional and research workflows) plus six Apache-2.0 Stanford CRFM HELM LegalBench tasks covering trademark distinctiveness, corporate lobbying, judicial opinion structure, international citizenship statutes, and private rights of action.

Weighting strict all-criteria pass rates alongside rubric partial credit distinguishes models that catch every indemnity cap and survival clause in an agreement from models that draft plausible prose while dropping one material exception.

## All rows are currently Estimated, and US-centric benchmarks do not replace counsel

Every one of the 40 ranked models carries an Estimated label (0 Supported) because independent public legal evaluations still cover fewer overlapping source families per model than general knowledge or coding suites. Treat sub-two-point gaps, such as the 0.2-point spread at the top, as practical ties.

Published legal benchmarks skew heavily toward US federal statutes, Delaware commercial contracts, and English-language case excerpts. They do not test local court rules, civil-law codes, privilege preservation, or citation shepardizing against current dockets; every output requires verification by qualified counsel in the governing jurisdiction.

[Inspect Harvey LAB & legal collection tables](/benchmarks/collections/legal) · [Claude Fable 5.2 vs Claude Opus 5.5](/compare/claude-fable-5-2-vs-claude-opus-5-5)

## Choosing models for contract review rooms versus high-volume clause extraction

Claude Fable 5.2 (86.1, $8/$40 per 1M tokens) and Claude Opus 5.5 (85.9, $4/$20 per 1M tokens) sit just 0.2 points apart at #1 and #2. Because Claude Opus 5.5 halves token cost to $4/$20 per 1M tokens with virtually identical legal benchmark accuracy, it is the stronger default for multi-agreement diligence rooms unless your own redline evals show a measurable edge for Claude Fable 5.2.

Law firms and corporate legal departments that prohibit third-party cloud processing can evaluate DeepSeek V4 Pro (#13, 67.2) on self-hosted hardware, pairing it with deterministic citation-lookup tools to close the 18.9-point gap to proprietary leaders.

[Documents & extraction leaderboard](/leaderboards/documents) · [Factuality & grounding leaderboard](/leaderboards/factuality)

## Questions

### Which LLM is best for legal work right now?

Claude Fable 5.2 holds #1 on our legal work leaderboard at 86.1 (Estimated) as of November 21, 2026, ahead of Claude Opus 5.5 at 85.9 and Claude Fable 5.1 at 81.1. Every ranked row currently carries an Estimated label, so validate top contenders on your own workload.

### What is the best open-weight LLM for legal work?

DeepSeek V4 Pro from DeepSeek leads open-weight models on this legal work board at #13 with a calibrated score of 67.2 (Estimated). That trails Claude Fable 5.2 by 18.9 points, so choose DeepSeek V4 Pro when self-hosting, VPC data residency, or weight customization matters more than peak frontier accuracy.

### Which API model offers the best value for legal work?

Claude Opus 5.5 (#2, 85.9) is our top API value pick for legal work, priced at $4/$20 per 1M tokens compared with Claude Fable 5.2 at $8/$40 per 1M tokens. You give up 0.2 calibrated points, making Claude Opus 5.5 the better buy whenever those extra points do not cut downstream retries.

### Which benchmarks make up the legal work leaderboard score?

BenchAlign v6.0 calibrates 11 open and primary domain protocols for this board, led by Legal Benchmarks: Contract Workflows (criteria) (16%), Legal Benchmarks: Contract Workflows (pass rate) (16%), DELTA Legal Research (criteria) (16%), Legal Benchmarks: Data Extraction (criteria) (16%). Each percentage reflects share of the vertical's reference weight inside our item-response model rather than a raw score average, and

### Why do all legal leaderboard models carry Estimated evidence labels?

All 40 models on the legal board currently hold Estimated status because public legal evaluations, primarily Legal Benchmarks contract, extraction, and DELTA research suites plus Stanford HELM LegalBench, come from fewer independent evaluator families per model than coding or math. Treat sub-two-point differences as statistical ties during vendor selection.

### What legal tasks are covered and left out by these benchmarks?

Scored tasks cover commercial contract redlining, structured clause extraction, US statutory research memos, and LegalBench doctrinal classification. They leave out non-US civil codes, local court filing procedures, privilege logs, and live citator checks for overruled precedent, so licensed counsel must verify every output.

## Related benchmarks and guides

- [Legal Benchmark Collection](/benchmarks/collections/legal): Inspect Harvey LAB, PRBench Legal, LegalBench, and case-research protocols.
- [Documents & Extraction Leaderboard](/leaderboards/documents): Long-PDF parsing and structured field extraction rankings.
- [Factuality Leaderboard](/leaderboards/factuality): Source-grounded verification and citation faithfulness rankings.
- [Best Long-Context LLMs](/best/long-context): Compare context windows for multi-agreement diligence rooms.

## Other specialist vertical leaderboards

- [Cybersecurity leaderboard](/leaderboards/cybersecurity) — #1 Gemini 4 Carbon (90.0)
- [Documents and extraction leaderboard](/leaderboards/documents) — #1 GPT-6.2 Astra (86.5)
- [Factuality and anti-hallucination leaderboard](/leaderboards/factuality) — #1 Gemini 4 Carbon (78.8)
- [Finance and accounting leaderboard](/leaderboards/finance) — #1 Gemini 4 Carbon (89.2)
- [Frontend and web development leaderboard](/leaderboards/frontend) — #1 Claude Fable 5.2 (89.5)
- [Healthcare and medical leaderboard](/leaderboards/healthcare) — #1 Claude Fable 5.2 (87.5)
- [Science and research leaderboard](/leaderboards/science) — #1 Claude Fable 5.2 (94.1)
- [Search and deep research leaderboard](/leaderboards/search) — #1 Claude Fable 5.2 (86.6)

Last updated: November 21, 2026

Canonical page: https://benchlm.ai/leaderboards/legal
