Legal AI benchmarks

Drafting, extraction, legal reasoning and US legal research. 3 public leaderboards below.

Legal Benchmarks leaderboard

Real lawyer requests (contract drafting, document extraction) run by an AI agent twice, then graded criterion by criterion by a judge calibrated against lawyers.

Run by Legal Benchmarks with Clavia. Leaderboard as of 9 Sep 2026 · retrieved 21 Oct 2026.

Anthropic
OpenAI
Google
xAI
Meta
DeepSeek
Alibaba
Z.ai
Moonshot
MiniMax
Mistral

Criterion Pass²: Share of substance criteria (legal correctness, completeness, reasoning) passed in both of 2 independent attempts.
Form Pass²: Share of form criteria (clarity, length, structure) passed in both attempts, averaged across tasks.
Cost per attempt (USD): Standard market token cost of one attempt, as published.

Legal Benchmarks leaderboard scores
Rank	Model	Provider	Criterion Pass²	Form Pass²	Cost per attempt (USD)
1	Claude Fable 5.1	Anthropic	90.4%	80.3%	$1.33
2	Claude Opus 5	Anthropic	89.7%	79.2%	$0.83
3	GPT-6 Astra	OpenAI	87.6%	87.8%	$1.03
4	Muse Spark 1.3	Meta	87.0%	77.8%	$0.24
5	GPT 5.6 Sol	OpenAI	83.7%	84.1%	$0.19
6	MiniMax M3	MiniMax	80.0%	78.2%	$0.028
7	Kimi K3	Moonshot	80.0%	79.9%	$0.37
8	GPT 5.6 Terra	OpenAI	80.0%	83.5%	$0.13
9	Claude Sonnet 5	Anthropic	79.6%	78.9%	$0.31
10	GPT 5.6 Luna	OpenAI	79.5%	82.9%	$0.022

Show 9 more rows

Rank	Model	Provider	Criterion Pass²	Form Pass²	Cost per attempt (USD)
11	DeepSeek V4 Pro	DeepSeek	70.2%	78.0%	$0.11
12	GLM-5.3	Z.ai	69.8%	74.7%	$0.10
13	Grok 4.6	xAI	68.5%	73.1%	$0.36
14	Gemini 3.8 Flash	Google	68.2%	71.6%	$0.22
15	Qwen 3.8 Flash	Alibaba	63.5%	66.1%	$0.029
16	GLM-5.3 Flash	Z.ai	52.3%	57.7%	$0.026
17	Qwen 3.8 Max	Alibaba	50.0%	63.2%	$0.94
18	DeepSeek V4 Flash	DeepSeek	49.2%	58.6%	$0.017
19	Mistral Medium 3.5	Mistral	45.9%	68.7%	$0.41

Source: Legal Benchmarks with Clavia leaderboard ↗

Top model here: Claude Fable 5.1 (Anthropic). Try prompts for Claude →

LegalBench

Open-source legal reasoning tasks such as issue spotting, rule recall and contract interpretation, scored on accuracy.

Run by Vals AI. Leaderboard as of 13 Oct 2026 · retrieved 21 Oct 2026. Showing top 25 of 151 models listed.

Anthropic
OpenAI
Google
xAI
Meta
Z.ai
Moonshot
MiniMax

Accuracy: Share of LegalBench questions answered correctly.

LegalBench scores
Rank	Model	Provider	Accuracy	Price in / out ($/1M tokens)
1	Claude Fable 5	Anthropic	88.56%	$10 / $50
2	Claude Fable 5.1	Anthropic	88.51%	$10 / $50
3	Gemini 4 Argon	Google	88.30%	$4 / $20
4	Wonder Full	Meta	87.42%	$4 / $20
5	Gemini 3.1 Pro Preview (02/26)	Google	87.40%	$2 / $12
6	Gemini 3.7 Flash	Google	87.26%	$1.5 / $7.5
7	Gemini 3 Pro (11/25)	Google	87.03%	$2 / $12
8	Gemini 3.8 Flash	Google	86.99%	$1.5 / $7.5
9	Claude Opus 5	Anthropic	86.97%	$5 / $25
10	GPT-5.6 Sol	OpenAI	86.97%	$4 / $20

Show 15 more rows

Rank	Model	Provider	Accuracy	Price in / out ($/1M tokens)
11	Gemini 3 Flash (12/25)	Google	86.86%	$0.5 / $3
12	Gemini 3.6 Flash	Google	86.70%	$1.5 / $7.5
13	GPT 5.5	OpenAI	86.52%	$5 / $30
14	Grok 4.6	xAI	86.31%	$2 / $6
15	GLM Legal	Z.ai	86.23%	$0.075 / $0.25
16	Kimi K3	Moonshot	86.21%	$3 / $15
17	GPT 5.4 (xhigh)	OpenAI	86.04%	$2.5 / $15
18	GPT 5	OpenAI	86.02%	$1.25 / $10
19	Grok 4.5	xAI	85.97%	$2 / $6
20	GPT 5.1	OpenAI	85.68%	$1.25 / $10
21	MiniMax-M3	MiniMax	85.42%	$0.6 / $2.4
22	Claude Opus 4.6 (Thinking)	Anthropic	85.30%	$5 / $25
23	Muse Spark 1.2	Meta	85.26%	$1.25 / $4.25
24	Claude Opus 4.7	Anthropic	85.25%	$5 / $25
25	GPT-5.6 Terra	OpenAI	85.11%	$2 / $12

Source: Vals AI leaderboard ↗

Top model here: Claude Fable 5 (Anthropic). Try prompts for Claude →

Legal Research Bench

AI agents research questions across areas of US law; a task only counts if every required point is right.

Run by Vals AI. Leaderboard as of 19 Oct 2026 · retrieved 21 Oct 2026. Showing top 25 of 82 models listed.

Anthropic
OpenAI
Google
xAI
Meta
DeepSeek
Alibaba
Z.ai
Moonshot
Xiaomi
Other

All-Pass: Share of research tasks where the answer met every grading point.
Cost per test (USD): Average API cost to run one task, as published.

Legal Research Bench scores
Rank	Model	Provider	All-Pass	Cost per test (USD)	Price in / out ($/1M tokens)
1	Muse Spark 1.3 Max	Meta	55.29%	$0.59	$1.25 / $4.25
2	Claude Opus 5	Anthropic	55.29%	$6.76	$5 / $25
3	Claude Fable 5.1	Anthropic	55.29%	$23	$10 / $50
4	Gemini 4 Argon	Google	54.81%	$6.45	$4 / $20
5	Wonder Full	Meta	53.37%	$3.60	$4 / $20
6	Wonder	Meta	50.48%	$3.98	$4 / $20
7	Claude Opus 5.5	Anthropic	50.48%	$25	$4 / $20
8	Claude Fable 5	Anthropic	49.52%	$9.79	$10 / $50
9	GLM 5.3	Z.ai	49.04%	$2.24	$1.4 / $4.4
10	Grok 4.6	xAI	48.08%	$1.53	$2 / $6

Show 15 more rows

Rank	Model	Provider	All-Pass	Cost per test (USD)	Price in / out ($/1M tokens)
11	GPT-5.6 Sol	OpenAI	48.08%	$22	$4 / $20
12	Claude Sonnet 5.5	Anthropic	48.08%	$15	$2 / $10
13	Qwen 3.8 Max	Alibaba	47.60%	$2.49	$2 / $6
14	Grok 4.7	xAI	47.12%	$4.92	$2 / $6
15	MiMo V2.6 Pro	Xiaomi	47.12%	$0.18	$0.435 / $0.87
16	Kimi K3	Moonshot	46.15%	$3.47	$3 / $15
17	DeepSeek V4.1 Flash 1017	DeepSeek	45.19%	$0.70	$0.3 / $1.2
18	GLM 5.3 Flash	Z.ai	45.19%	$0.12	$0.075 / $0.25
19	Hy4 Preview	Other	45.19%	$0.85	$0.834 / $2.501
20	Ember-1	Other	44.71%	$3.67	$3 / $15
21	Muse Spark 1.2	Meta	43.75%	$0.51	$1.25 / $4.25
22	Claude Opus 4.8	Anthropic	43.75%	$2.82	$5 / $25
23	Claude Haiku 5.5	Anthropic	43.27%	$1.63	$0.1 / $0.5
24	Claude Sonnet 5	Anthropic	41.83%	$2.72	$2 / $10
25	GLM Legal	Z.ai	41.83%	$0.050	$0.075 / $0.25

Source: Vals AI leaderboard ↗

Top model here: Muse Spark 1.3 Max (Meta). Try prompts for Muse →

ChatJudge does not run these tests. Scores are copied as published by each source, with a link and the date we read them, and only change when we refresh from the source. Different leaderboards use different settings, so compare models within one chart, not across charts. Spotted a mistake? Tell us.