← Back to HAQQ Blog

Best AI for Legal Work in 2026: October Update + 3,000 Graded Answers

By Stephane Boghossian · · Updated · 15 min read · Ai-legal-tech

October 2026: Claude Opus 5.5 and GPT-6 Astra lead our latest legal tests. Plus 3,000 graded answers from 10 models, 24% of which cite law that does not back them.

Key facts

Three runs in September 2026 replaced the June leaderboard for the frontier-model question. Each tests something different, so read them together.

ModelVendorRunRun dateResult
Claude Opus 5.5AnthropicFive MENA trap questions30 Sep 20263.30 / 4, named 0 invented judgments
GPT-6 AstraOpenAIFive MENA trap questions30 Sep 20263.00 / 4, named 0 invented judgments
DeepSeek V4 ProDeepSeekFive MENA trap questions30 Sep 20262.20 / 4, named 8 unverifiable judgments
Gemini 3.1 ProGoogleFive MENA trap questions30 Sep 20262.10 / 4, named 8 unverifiable judgments
Claude Opus 5AnthropicTen drafting and advice prompts, pairwise23 Sep 202639 of 80 verdicts, $0.4362 per answer
Claude Fable 5.1AnthropicTen drafting and advice prompts, pairwise23 Sep 202636 of 80 verdicts, $0.7249 per answer
Claude Opus 5.5AnthropicTen drafting and advice prompts, pairwise23 Sep 202635 of 80 verdicts, $0.3647 per answer
GPT-6 AstraOpenAI41 prompts, 20 practice areas10 Sep 202617.63 / 20 substance, $0.2622 per answer
GPT-5.6 Luna ProOpenAI41 prompts, 20 practice areas10 Sep 202630.93 / 35 composite (with speed), $0.0231 per answer
Claude Opus 5Anthropic41 prompts, 20 practice areas10 Sep 202628.13 / 35 composite, $0.4566 per answer
Claude Opus 4.8Anthropic300 commercial tasks (superseded)Jun 202630.02 / 35, won 130 of 300 tasks

October 2026 update: which result is current

What the September runs agree on: the quality gap between current frontier models on legal tasks is small, and the useful differences are in how a model fails (inventing authority, misreading a local rule) and what it costs per answer. Test the shortlist on your own matters before you pick.

On the measure that matters most in research, whether the authority exists and says what the model claims, Claude Opus 5.5 and GPT-6 Astra behaved best in our latest run: zero invented judgments in the eight answers between them. On Vals AI's independent Legal Research Bench, Opus 5.5 scored 50.48%, 4th of 68 models when we read it on 23 September 2026. No model we have tested is accurate enough to cite without checking. In September, all three models in the Astra run gave the right article number attached to the wrong rule on a Saudi or Egyptian question, which a citation checker would pass.

That is the gap a research tool has to close above the model: retrieve the primary source, check the citation against it, and say when it cannot find one. It is how we built HAQQ legal research.

Does a bigger or newer model hallucinate less?

Newer models hallucinate less often, and they still do it. Between June and September the pattern changed: in June, every one of the 10 models fabricated or misapplied a citation somewhere in 300 tasks, and 24% of all answers had a citation problem. In September, the newest Anthropic and OpenAI flagships declined to invent case law when asked for it. Gemini 3.1 Pro and DeepSeek V4 Pro still produced case numbers on demand, a different pair each time we asked. Model size alone did not predict the behaviour; the two that refused were also the two most expensive per answer in that run.

Earlier run: NY litigation strategy (Jan 2026)

In January 2026 we gave seven systems one prompt: "Prepare litigation strategy under New York law for unpaid invoice of $250,000." We scored each answer out of 10 on six criteria. Strong answers used CPLR 3213 (summary judgment in lieu of complaint), pleaded breach of contract and account stated, and planned post-judgment collection (restraining notices, bank levies, turnover orders). Weaker answers stopped at filing suit. It was a single prompt scored by HAQQ, with HAQQ among the entrants, and every general model in it has since been replaced, so read it as a record of how answers differed, not as a ranking.

ModelLegal accuracyProcedural depthStrategyAuthoritiesStructureClient-ready
HAQQ9.59.5999.59.5
Claude Opus 4.698.59998.5
GPT-5.28.58.5978.58.5
Gemini 3.1 Pro8887.588
Grok 4.17.57777.57
Mistral Large 3776.5676.5
Perplexity Sonar66666.56

How did we run the June 300-task benchmark?

We wrote 300 original legal tasks with named parties, amounts, dates and governing statutes: draft this clause, redline this provision, structure this transaction, analyse this conflict of laws. They span 51 practice areas and more than 20 jurisdictions, including US federal law and Delaware, California, New York and Texas, the UK, the EU, the UAE and DIFC, Saudi Arabia, Lebanon, Egypt, Qatar, Singapore, Australia, Canada, India, Brazil, Nigeria, OHADA, Japan, Germany, France and Switzerland, plus the Hague Conventions and Cayman, BVI and Bermuda.

Difficulty was weighted hard: 114 tasks at level 5 of 5, 108 at level 4, 22 at level 3. Every task went to all 10 models with an identical system prompt at temperature 0, capped at 6,000 output tokens. Each answer was scored on five dimensions:

Claude Sonnet 4.6 scored quality, accuracy, style and creativity against a fixed rubric; speed came from real latency. The caveats section covers judge bias.

What did the June 2026 leaderboard show?

How often do AI models cite law that does not support the answer?

In June, 24% of all 3,000 answers cited or applied law that does not say what the model claimed: invented cases, misapplied statutes, the right doctrine attached to the wrong jurisdiction. The judge flagged specific, checkable errors. One example per model:

Every model, including the leaders, made at least one such error, and the top accuracy score was 8.41 of 10 (GPT-5.5). The research tools lawyers already pay for show the same problem: Stanford RegLab's 2024 study measured hallucinations in about a third of Westlaw AI-Assisted Research answers and more than one in six for Lexis+ AI.

Does one model win every practice area?

No. Across the 51 practice areas in June, Claude Opus 4.8 led 30, Grok 4.3 led 13, o3 led 6, and Gemini 3.1 Pro and Claude Sonnet 4.6 led one each.

The September Astra run found the same shape at a smaller scale: Astra ahead in most practice areas, Claude Opus 5 ahead in insolvency and ESG. A legal product that routes each matter to the model most likely to get it right, then checks the answer, gets more out of the field than one that bets on a single model.

In June, cost per task spanned a 90x range, from $0.0009 (DeepSeek V3.2) to $0.082 (GPT-5.5), and latency ran from 7.7 seconds (Llama 4 Maverick) to 134 seconds (GPT-5.5). The September runs, priced at OpenRouter list prices on the run date, give the current picture:

For a high-volume product, the model with the top quality score can be the wrong choice once cost and latency are counted. Astra's raised reasoning effort is a case in point: in September, moving it from low to high multiplied cost by 1.5x and latency by 1.8x and moved judged quality by minus 0.17 points.

We keep a second leaderboard: a 50-task evaluation on a 50-point rubric across 11 task categories that adds the legal platforms firms shortlist (Harvey, CoCounsel, Legora, Spellbook, LexisNexis +AI, Clio Duo) next to frontier models. It is the leaderboard on our compare page, and HAQQ ranks first on it (see the disclosure above). Ranked by average score:

RankModel or productAverage /50
1HAQQ (Justinian)46.8
2Claude Fable 542.7
3Claude Opus 4.840.8
4Mike OS39.1
5Harvey Tenet37.9
6DeepSeek V4 Pro36.8
7Thomson Reuters CoCounsel36.6
8Legora35.9
9GPT-5.6 Sol Pro35.3
10Claude + legal plugins33.9
11Gemini 3.1 Pro32.9
11Spellbook32.9
13LexisNexis +AI32.0
14Grok 4.530.7
15Perplexity Sonar27.2
16Clio Duo25.6
17Llama 4 Maverick23.8
18Mistral 322.4
19Qwen3.7 Plus18.7

Two readings stand out. The legal platforms cluster in the middle: Harvey Tenet (37.9), CoCounsel (36.6), Legora (35.9) and LexisNexis +AI (32.0) all sit below the strongest frontier models on standalone answer quality, so what a firm pays a platform for is mainly workflow, integrations and security. And the spread is wide, from 46.8 to 18.7, so the choice of tool changes the output a lawyer starts from. A per-task score cannot see firm-wide deployment, certifications or document-scale workflows, which are real strengths of the larger platforms.

The platform leaderboard has not yet been re-run on Opus 5.5, Fable 5.1 or GPT-6 Astra; the September runs above are the current source for those models.

Which provider leads across every run?

Collapsing models to provider and merging the June run with every earlier run of the same commercial benchmark (0 to 100 index, weighted by number of evaluations):

The top providers sit close together. What you build around the model (retrieval, citation checks, jurisdiction rules) moves the result more than which of them you pick.

Can I trust these benchmarks?

How should a firm choose a model today?

Sources

FAQ

What is the best AI for legal work in 2026?

As of October 2026, Claude Opus 5.5 and GPT-6 Astra are the strongest frontier models in our tests. On a five-question UAE, Saudi and Lebanese trap test run on 30 September 2026 they scored 3.30 and 3.00 of 4, ahead of DeepSeek V4 Pro (2.20) and Gemini 3.1 Pro (2.10), and neither invented a court judgment when asked for one. The gap between the two is within run-to-run noise, so price and your own matters should decide.

Which AI is most accurate for legal research?

In our 30 September 2026 run, Claude Opus 5.5 and GPT-6 Astra declined to name Dubai Court of Cassation judgments for an article that does not exist, while Gemini 3.1 Pro and DeepSeek V4 Pro named 16 that did not match any public source as cited. Vals AI's independent Legal Research Bench placed Opus 5.5 4th of 68 models at 50.48% (read 23 September 2026). No model is accurate enough to cite without checking the source.

Does a bigger or newer AI model hallucinate less on legal questions?

Newer flagships hallucinate less often but still do. In our June 2026 run of 3,000 answers, 24% cited or applied law that did not support the claim, and every one of the 10 models erred somewhere. In September 2026 the newest Anthropic and OpenAI models refused to invent case law, but all three models in our GPT-6 Astra run misstated one Saudi or Egyptian rule while citing the correct article.

Is Claude Opus 5.5 better than Claude Fable 5.1 for legal work?

We could not tell them apart on quality. Across 120 blind verdicts from two non-Anthropic judges on 23 September 2026, Opus 5.5, Opus 5 and Fable 5.1 finished within four verdicts of each other and formed a loop rather than a ranking. Opus 5.5 cost $0.3647 per answer against $0.7249 for Fable 5.1 and was the fastest of the three.

How much does it cost to run a legal question through a frontier model?

In our September 2026 runs at OpenRouter list prices, GPT-5.6 Luna Pro cost about $0.02 per answer, GPT-6 Astra $0.26, Claude Opus 5.5 $0.36, Claude Opus 5 $0.44 to $0.46 and Claude Fable 5.1 $0.72. Answer length drives the bill as much as the per-token price: Astra charges twice Opus 5's rate per token and still costs 1.7x less per answer because it writes about 3.5x fewer tokens.

Where do legal AI platforms like Harvey rank against frontier models?

On HAQQ's 50-point task benchmark across 11 categories and 19 entrants, Harvey Tenet averages 37.9/50, ahead of CoCounsel (36.6), Legora (35.9) and LexisNexis +AI (32.0), and behind Claude Fable 5 (42.7) and Claude Opus 4.8 (40.8). Harvey's best placing in any single category is second. HAQQ authored that benchmark and ranks first on it at 46.8/50, so treat it as a vendor's leaderboard and check the method.

Which Claude model is best for legal work?

On HAQQ's 50-point legal benchmark (scores as of 20 October 2026), Claude Fable 5 averaged 42.7/50 and Claude Opus 4.8 40.8, both ahead of Harvey (37.9). Claude with legal plugins scored 33.9, below plain Claude. Plain Claude suits analysis, first drafts and long documents, but it is not a system of record or a citation checker. HAQQ ran the benchmark and ranks first in it, so re-test on your own matters.

Where can I find independent benchmark results for legal AI tools?

Stanford RegLab researchers (Magesh et al., 2024) independently measured hallucination rates of about 17% for Lexis+ AI and 33% for Westlaw AI-Assisted Research. HAQQ's 50-point benchmark of 19 tools and models publishes every category score on haqq.ai/compare-us, but it is not independent: HAQQ ran it, finishes first in it, and no third party has audited it.