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

Key facts

• On a five-question UAE, Saudi and Lebanese trap test, Claude Opus 5.5 scored 3.30 of 4 and GPT-6 Astra 3.00, against 2.20 for DeepSeek V4 Pro and 2.10 for Gemini 3.1 Pro, with hints removed (HAQQ run, 30 Sep 2026, 80 answers, three blind judges).
• Asked for Dubai Court of Cassation judgments supporting a non-existent article, Gemini 3.1 Pro and DeepSeek V4 Pro named 16 across eight answers; none matched a public source as cited (HAQQ run, 30 Sep 2026).
• Claude Opus 5.5, Opus 5 and Fable 5.1 finished within four verdicts of each other across 120 blind pairwise verdicts; Opus 5.5 cost $0.3647 per answer against $0.7249 for Fable 5.1 (HAQQ run, OpenRouter list prices, 23 Sep 2026).
• GPT-6 Astra led legal substance at 17.63 of 20 against Claude Opus 5 and GPT-5.6 Luna Pro, and all three got one Saudi or Egyptian rule wrong while citing the right article (HAQQ run, 41 prompts, 10 Sep 2026).
• Vals AI's Legal Research Bench placed Claude Opus 5.5 at 50.48%, 4th of 68 models (Vals AI, read 23 Sep 2026).
• In our largest run, 24% of 3,000 answers from 10 frontier models cited or applied law that does not say what the model claimed (HAQQ 300-task run, published 5 Jun 2026).
• Stanford RegLab measured hallucinations in about 33% of Westlaw AI-Assisted Research answers and more than 17% of Lexis+ AI answers (Magesh et al., 2024).

Which AI is best for legal work as of October 2026?

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

October 2026 update: which result is current

• Current for frontier models: the 30 September trap test (how to test legal AI). It is the most recent run, and it covers four current flagship models from four labs. Every model caught every trap. The split came on authority: Opus 5.5 and Astra declined to name judgments that could not be verified, while Gemini 3.1 Pro and DeepSeek V4 Pro produced 16 case citations, two of them real case numbers with the wrong date and holding, the other 14 not located. Opus 5.5 and Astra are the top pair, and the gap between them is within run-to-run noise.
• Current for choosing between Anthropic models: the 23 September run (Claude Opus 5.5 legal benchmark). Opus 5.5, Opus 5 and Fable 5.1 formed a loop (each beat one and lost to one), so quality did not separate them at ten prompts. Price did: Opus 5.5 answered at roughly half Fable 5.1's cost and was the fastest of the three.
• Current for GPT-6 Astra against cheaper models: the 10 September run (GPT-6 Astra on legal work). Astra led on substance by less than a point and cost 11 times more per answer than GPT-5.6 Luna Pro. Each of the three models misstated one Saudi or Egyptian rule while citing the correct article.
• Superseded: the June 300-task run below. It is still our largest data set and the source of the 24% citation-gap figure, but every model in it has a newer version. Use it for patterns (citation errors at every tier, no single model winning every practice area, a 90x cost spread), not for picking a model today.

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.

Which AI is most accurate for legal research?

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.

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:

• Quality (1 to 10): depth, completeness, actionability.
• Accuracy (1 to 10): legal correctness, minus 5 for hallucinated case law and minus 3 for the wrong jurisdiction.
• Speed (1 to 5): measured latency.
• Style (1 to 5): professional structure.
• Creativity (1 to 5): non-obvious risks and cross-jurisdiction issues spotted.

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?

• Claude Opus 4.8 placed first in 130 of 300 tasks (nearly double any other model), top three in 265, with the highest quality (8.9) and creativity scores. It was slow (60.8 s) and among the most expensive ($0.069 per task).
• Grok 4.3 placed second (28.98) in 8.8 seconds at $0.003 per task, the strongest result for speed and cost, and won the most environmental/ESG, IP and edge-case tasks.
• GPT-5.5 posted the highest accuracy (8.41) and the lowest hallucinated-citation rate (3%), yet placed fifth on total and won one task outright. It was the slowest (134 s) and the most expensive ($0.082).
• o3 won 66 tasks outright and still ranked eighth, with a 32% hallucination rate and accuracy of 5.89.
• Mistral Large misapplied or invented citations in 64% of answers; Llama 4 Maverick came last (20.01).

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:

• Claude Opus 4.8: Computer Associates / Gemstar cited as gun-jumping precedents; not accurate characterizations.
• GPT-5.5: Hitz citation appears fabricated; 616 B.R. 374 unverifiable.
• Gemini 3.1 Pro: Halpin v. Riverstone citation unverifiable; likely hallucinated Delaware case.
• Grok 4.3: CBS v. Ziff-Davis cited inaccurately; case concerns warranty reliance, not disclosure.
• Claude Sonnet 4.6: Schrier cite unverifiable; Biotronik case real but misapplied.
• o3: MSA Technology v Antec and GB Gas citations appear fabricated or misrepresented.
• Qwen3.7 Max: Specht v. Netscape misapplied; not a goods/services UCC case.
• DeepSeek V3.2: Crosstown Music citation unverifiable / misapplied.
• Mistral Large: N.Y. Gen. Oblig. Law §5-328 and UCC §2-515 citations are dubious / misapplied.
• Llama 4 Maverick: Cavendish case misapplied; unrelated to pricing clauses.

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.

• Opus 4.8 led the doctrinal core: M&A (30.6), regulatory compliance (30.5), securities, white collar, tax, banking, data privacy, employment and international trade.
• Grok 4.3 led IP and technology (30.2), environmental and ESG (30.8), consumer protection, and compliance and due diligence (31.4).
• o3 led legal operations (31.1), securities and finance (30.9), government contracts and procurement (31.1) and cross-border real estate.
• Gemini 3.1 Pro posted the top score in privacy and data protection (31.4).

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.

How much does each model cost per legal task?

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:

• Claude Opus 5.5: $0.3647 per answer at high reasoning effort, against $0.7249 for Fable 5.1 and $0.4362 for Opus 5 (23 Sep 2026).
• GPT-6 Astra: $0.2622 per answer, 1.7x cheaper than Claude Opus 5 because it writes about 3.5x fewer tokens, and 11x more than GPT-5.6 Luna Pro at $0.0231 (10 Sep 2026).
• On the trap test at default settings, the 20 answers per model cost USD 4.04 for Opus 5.5, 2.81 for Astra, 0.92 for Gemini 3.1 Pro and 0.88 for DeepSeek V4 Pro (30 Sep 2026).

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.

How do legal AI platforms compare with frontier models?

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:

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?

• Output caps. Our first June pass capped answers at 1,200 tokens, which starved reasoning models and clipped verbose ones. We discarded it and re-ran at 6,000. Most models gained 1.5 to 2.5 points; Mistral and Llama got worse, because longer answers exposed more bad citations. In September a 32,000-token cap truncated two Opus answers on one prompt; we re-ran at 64,000 and now abort any run with a truncated answer.
• Judges. June used one judge, Claude Sonnet 4.6, so self-preference is a fair question; Grok 4.3 placing second and GPT-5.5 posting the top accuracy argue against a strong home bias, but a second judge would be better. The September runs used two non-Anthropic judges (23 Sep) and three judges from three model families with the median taken (30 Sep). On 23 September the two judges picked different winners from the same answers.
• Sample size. The September runs used 10, 41 and 5 prompts. That is enough to see a large difference and too small to rank models that are close. Read effect sizes, not decimals.
• Single runs. June was a single run at temperature 0. In September, repeated answers moved by 0.3 points on average and by a full point or more in 11 of 40 pairs.
• Human review. No practising lawyer graded the September answers, and Arabic drafting was judged by models only.

How should a firm choose a model today?

• Shortlist Claude Opus 5.5 and GPT-6 Astra for research and advice where invented authority is the main risk; both declined to fabricate judgments in our latest test.
• Price the shortlist on your own volume. Opus 5.5 is about half the cost of Fable 5.1 per answer for quality we could not tell apart; GPT-5.6 Luna Pro is far cheaper than Astra for general Q&A where a point of quality is not worth 11x.
• Test on your own jurisdictions. Every model in the Astra run misread one Saudi or Egyptian rule while citing the right article.
• Ask for authority that cannot exist, ask twice, and check every citation by hand. The full method is in how to test legal AI.
• Re-check every quarter. The June leader has three successors on this page already.

Sources

• HAQQ, Claude Opus 5.5 legal benchmark (23 Sep 2026)
• HAQQ, GPT-6 Astra on legal work (10 Sep 2026 run)
• HAQQ, How to test a legal AI: five trap questions (30 Sep 2026 run)
• Vals AI, Legal Research Bench
• Anthropic, Introducing Claude Opus 5.5
• Magesh et al., Stanford RegLab study of AI legal research tools (2024)

Related reading

• the court-filing hallucination audit
• nine AI systems on a Lebanese waqf matter
• when an LLM judge rewards fabrication
• the scope and method of our MENA legal AI benchmark
• why LLMs fabricate, explained for lawyers