← Back to HAQQ Blog

Best LLM for Legal Writing in 2026: Opus 5.5, Fable 5.1, GPT-6 Astra, Gemini

By Stephane Boghossian · · Updated · 8 min read · Ai-legal-tech

Which model should write your client alert, law review piece or firm blog? Drafting scores, cost per answer and citation behaviour from our September 2026 legal runs.

Key facts

What does an LLM need to do to write a legal article?

A legal article goes out under a lawyer's or a firm's name to readers who may act on it. The model has to do five things, and our benchmarks test most of them directly.

No single run tested all four models on the same prompts, so the table shows each model's figures from the run that included it. The runs used different settings (high reasoning effort on 23 September, medium on 10 September, defaults on 30 September), so compare costs within a column, not across it.

ModelDrafting score (23 Sep, /10)Legal substance (10 Sep, /20)Trap test (30 Sep, /4)Cost per answerCitation behaviour
Claude Opus 5.58.56not tested3.30$0.3647 (23 Sep)Named no invented judgments
Claude Fable 5.18.70not testednot tested$0.7249 (23 Sep)Accuracy 8.55, tied top
GPT-6 Astranot tested17.633.00$0.2622 (10 Sep)Named no invented judgments; misstated one Egyptian data-law rule while citing the right provision
Gemini 3.1 Pronot testednot tested2.10about $0.046 (30 Sep, $0.92 for 20 answers)Named 8 judgments, none matched as cited

We do not publish a cost per 1,000 words: reasoning tokens bill as output and are not part of the visible text, so the per-answer cost from a real run is the more useful figure. Gemini 3.1 Pro's low cost comes partly from shorter thinking: 3,294 median output tokens per answer in the 30 September run, against 10,124 for Opus 5.5.

Which AI should write a client alert?

GPT-6 Astra. A client alert needs to be short, correct and fast to review. In our 10 September run Astra led on legal substance (17.63 of 20) and wrote about 3.5x fewer tokens per answer than Claude Opus 5, which is why it cost 1.7x less per answer despite twice the per-token price. In the 30 September trap test it declined to invent case law. The caveat that matters for alerts in our region: Astra misstated one rule of Egyptian data protection law while citing the correct provision, so a lawyer still checks the substance of every rule the alert relies on.

If cost per alert matters more than the last point of quality, GPT-5.6 Luna Pro beat Astra on the composite score in the same run (30.93 against 30.05 of 35, a score that includes speed), trailed it on substance by less than a point, and cost $0.0231 per answer, 11x less.

Which AI should draft a law review article?

Claude Fable 5.1, with a small margin. Long-form academic writing rewards drafting quality and accuracy together, and Fable 5.1 had the highest drafting score (8.70) and tied for the top accuracy score (8.55) in our 23 September run. The margin over Opus 5.5 (8.56) is inside the noise of a ten-prompt run, and the two judges disagreed on which model wrote better overall: GPT-5.5 as judge preferred Fable 5.1, Gemini 3.1 Pro as judge preferred Claude Opus 5. Pay Fable's roughly double price per answer for a piece where drafting is the product, and expect to verify every footnote yourself.

Which AI should write a law firm blog post?

Claude Opus 5.5. Firm blogs run at volume, so price and speed count. On 23 September Opus 5.5 tied Fable 5.1 on accuracy (8.55), sat within 0.14 points of it on drafting, answered at about half the cost per answer ($0.3647 against $0.7249) and was the fastest of the three Anthropic models (184 seconds per answer on average). It also scored highest in our 30 September trap test (3.30 of 4) and declined to invent judgments. We moved our own routine legal drafting to Opus 5.5 after that run.

For outlines, summaries of material you supply, and first drafts you will rewrite, yes: it was the fastest model in our 30 September run (27 seconds median per answer) and the cheapest after DeepSeek V4 Pro. For anything that cites authority, treat its citations as unverified. Asked for two Dubai Court of Cassation judgments supporting a non-existent article, it named two in every answer, a different pair each run, and none of the eight matched a public source as cited. It also scored lowest of the four models on that test (2.10 of 4 with hints removed).

Custom GPTs in OpenAI's GPT Store with names like 'LegalGPT' are instructions and files layered on the underlying OpenAI model, so their writing quality follows the base model they run on. We have not benchmarked them individually. 'Claude + legal plugins' is an entrant in HAQQ's 50-point task benchmark; it scored below Claude Opus used on its own in that benchmark, which suggests a plugin's prompts can constrain a strong model as easily as improve it. Test any wrapper on your own article before you trust it over the base model.

What prompts work for each type of legal article?

Client alert

Write a client alert of no more than 400 words on [law or regulation, with number and date].
Jurisdiction: [explicit].
Audience: general counsel and CFOs.
Include: what changed, the effective date, who is affected, what remains unclear, and two questions the reader should ask their adviser.
Cite the article for every rule you state. Use only the attached text of the law. Do not recommend actions to a specific client.

Law review or long-form commentary

Write a 3,000-word commentary on [question of law] for legal academics and practitioners.
Jurisdiction: [explicit].
Set out at least two competing interpretations, the authorities for each, and your reasoned position.
State your assumptions and the limits of the analysis.
Footnote every authority in [citation style], using only the sources I attach. Where no attached source supports a point, write [AUTHORITY NEEDED] instead of citing one.

Firm blog post

Write a 900-word blog post for [audience, e.g. founders in the UAE] explaining [topic].
Jurisdiction: [explicit].
Structure: the question in the reader's words, the short answer, how the rule works, common mistakes, when to get advice.
Plain English, no case citations in the body, a one-line general-information disclaimer at the end.
Flag any statement you are not certain is current law.

More templates, including drafting and client-communication prompts in five languages, are in our prompting guide for lawyers.

How should a lawyer check an AI-written legal article?

The checks above take longer than the drafting. That is where a legal platform earns its place: in HAQQ Legal AI Chat, drafts are grounded in sources retrieved before the model answers and every citation is traceable to its source, so the check starts from the source instead of a search for it.

Sources

FAQ

What is the best LLM for legal writing in 2026?

It depends on the article. In HAQQ's September 2026 runs, GPT-6 Astra led on legal substance (17.63 of 20) with short answers, which suits client alerts; Claude Fable 5.1 had the top drafting score (8.70 of 10), which suits long-form commentary; and Claude Opus 5.5 matched Fable on accuracy at about half the cost per answer, which suits firm blogs at volume. The margins are small, so test on your own articles.

Is Claude Opus 5.5 or Claude Fable 5.1 better for legal drafting?

We could not separate them on quality. On 23 September 2026, across ten lawyer-grade prompts and two non-Anthropic judges, Fable 5.1 scored 8.70 on drafting and Opus 5.5 8.56, and both scored 8.55 on accuracy. Opus 5.5 cost $0.3647 per answer against $0.7249 for Fable 5.1 and answered faster.

Is GPT-6 Astra good for writing legal content?

For short, substantive pieces, yes. In HAQQ's 10 September 2026 run it scored 17.63 of 20 on legal substance and wrote about 3.5x fewer tokens per answer than Claude Opus 5, and on 30 September it declined to invent case law when asked for it. It still misstated one Egyptian data protection rule while citing the right provision, so a lawyer must check each rule it states.

Can Gemini 3.1 Pro write a legal article?

It can draft outlines and summaries quickly and cheaply, but do not rely on its citations. In HAQQ's 30 September 2026 test it named eight Dubai Court of Cassation judgments for an article that does not exist, none of which matched a public source as cited, and it scored lowest of four models (2.10 of 4).

How do I stop an AI from inventing citations in a legal article?

Give it the sources and tell it to cite only those, with a placeholder such as [AUTHORITY NEEDED] where nothing supports a point. Then open every citation and confirm it says what the draft claims, check the date of each rule, and ask the same citation question twice: a real authority comes back the same, a generated one often changes.

Which Claude model is best for legal writing?

On HAQQ's 50-point legal benchmark, Claude Fable 5 scored 45/50 on contract drafting and 41/50 on memo drafting, against 43 and 41 for Claude Opus 4.8, and averaged 42.7 to 40.8 across all 11 categories. HAQQ built the benchmark and ranks first in it, so test both on your own drafts before choosing.

Is Claude with a legal plugin better than Claude on its own?

No, on HAQQ's 50-point task benchmark. Claude + legal plugins averaged 33.9/50 against 40.8 for Claude Opus 4.8 used on its own, which outscored it in 11 of 11 categories. A plugin's prompts can constrain a strong model as easily as improve it. HAQQ built the benchmark, so treat it as vendor testing and test any wrapper on your own work.