AI Agents for Lawyers

Guides

Deep Research Tools for Lawyers: Claude, ChatGPT, Gemini and Perplexity Compared

Deep research AI for lawyers compared: what Claude, ChatGPT, Gemini and Perplexity research modes do, where they fail on law, and a verification table.

By Dr. Niklas SchmidtUpdated 13 November 202610 min read
On this page
  1. What a deep-research run does that a chat does not
  2. The four tools side by side
  3. Where they fail on law
  4. A prompt that gets the most out of any of them
  5. Grounded alternatives and what "grounded" buys
  6. Verifying a deep-research report before you rely on it
  7. A setup that works

A deep-research mode is the fastest way to get a sourced 20-page briefing on an unfamiliar area of law and the worst way to find a citation you intend to file. Claude, ChatGPT, Gemini and Perplexity all now run multi-step agents that plan, read dozens of pages and write up with footnotes; a University of Pennsylvania study in April 2026 found they fabricate 3–13% of the URLs they cite, more than a plain chat, because they cite more.

What a deep-research run does that a chat does not

A standard chat answers from the model's training and, at best, a handful of search results. A deep-research run is a small agent: it turns your question into a plan, issues its own searches, reads the pages it finds, revises the plan, and writes a structured report with sources. Reported run times differ by an order of magnitude, as of 2026: ChatGPT's deep research runs up to 30 minutes, Perplexity's two to four, and Gemini's reads 100 or more pages per run (Superkind collated the tier limits; its figures are a single secondary source and worth checking against the vendor pages before you rely on them).

Three consequences follow for legal use. The report is only as good as what the agent could reach, and most primary law sits behind Westlaw, Lexis, Beck-online or RIS paywalls the agent cannot enter. The report cites more than a chat does, which is why the fabrication rate rises rather than falls. And the report reads as finished work, which is precisely the property that makes lawyers stop checking; Scientific American's May 2026 review of hallucination sanctions calls this "cognitive surrender". The hallucinations guide covers the mechanism; this page is about using the modes anyway.

The four tools side by side

As of September 2026, on the things a lawyer should ask before the first query:

ClaudeChatGPTGeminiPerplexity
Current research modelOpus 5 (24 Jul 2026), Fable 5.1 (1 Sep 2026, paid tiers only)GPT-6 Astra (public 4 Sep 2026; off by default in Enterprise until an admin enables it)Gemini 3.1 Pro (19 Feb 2026); 3.5 Pro announced 19 May 2026, unreleasedRoutes to 20+ models per sub-task; Computer for Counsel (24 Jun 2026)
Context and cutoff1M tokens; Opus 5 cutoff May 2026, Fable 5.1 June 2026~1.05M tokens; cutoff 30 April 2026; "recurrent depth" hides some reasoning1M tokens inputDepends on routed model
Legal reachClaude for Legal plugins tag unverified citations [verify]; CoCounsel partner plugin needs a Thomson Reuters subscriptionAstra for Law (17 Sep 2026): Legal Search Index of 230M+ URLs of US law built with CourtListener; 26 partner pluginsGemini Notebook (ex-NotebookLM, renamed 16 Jul 2026) is a closed-universe tool: 50 sources free, 300 paid400+ MCP tools; Midpage, Clio Vincent and DMS connectors announced
Training on your inputsNever on Team, Enterprise and API; consumer users choose (opted-in data kept 5 years)Consumer Free/Plus/Pro train unless you opt out; Business, Enterprise and API do notConsumer activity kept 18 months by default, a subset read by human reviewers, used for training unless "Keep Activity" is off; Workspace excludedEnterprise never trains by default; personal accounts have an "AI data retention" toggle
Zero data retentionAPI only, per organisation; Fable 5.1 requires 30-day retention and is not ZDR-eligible without express authorisation; Team and Enterprise chat not ZDR-eligibleContractual on API; Astra for Law offers a ZDR APIEnterprise data residency for Gemini Notebook in the EU since 3 Jun 2026Not published for chat
Entry pricePro $20; Team $25 standard / $125 premium seatPlus $20; Business $25 a seatAI Pro personal-account only; Workspace needs add-onsPro $20; Enterprise Pro $40 a seat
Independent legal benchmarkVals Index (15 Sep 2026): Fable 5.1 first, Opus 5 second; Legal Research Bench: Opus 5 second, Fable 5.1 thirdVals Index: Astra third; Astra for Law 54% on Legal Research Bench vs 38.7% plain web search (OpenAI-reported)LegalBench 87.40% (16 Sep 2026)None published for the routed system

Two rows decide most firm choices. The training row rules out any consumer tier for matter work, and the privacy settings walkthrough shows the toggles for anyone who ignores that advice. The ZDR row is the counter-intuitive one: Anthropic's strongest model, Fable 5.1, is the one that cannot run under zero data retention, so a firm that needs ZDR is choosing Opus 5 or Sonnet 5 on the API, not the top of the leaderboard (Anthropic retention docs). The Perplexity, Gemini Notebook and Claude tool pages go deeper on each.

Where they fail on law

The failure modes are specific and repeatable, which is good news: you can design around them.

Paywalled sources. A browsing agent reaches CourtListener, BAILII, EUR-Lex, RIS and the judiciary's own sites; it does not reach Westlaw, Lexis, Practical Law or Beck-online. For US case law that gap has narrowed since Astra for Law's 230-million-URL index built with the Free Law Project, launched on 17 September 2026 (LawSites); treat it as days old. For Austrian or German commentary literature the gap remains total.

Jurisdictional confusion. Vals' October 2025 study of 199 US research questions found every system, general and legal, lost roughly 11–14 points of weighted accuracy on multi-jurisdiction questions (LawSites summary). The 2025 "Place Matters" study, using older models, measured hallucination at 45% on Los Angeles questions, 55% on London, 61% on Sydney and 100% on Australian residential-tenancies law (AI Law Librarians). The further from US federal law, the worse.

Dated law. The model's cutoff is not the report's date. A run on GPT-6 Astra knows nothing natively after 30 April 2026; the browsing fills some of the gap but cannot know what it did not search for. A report on the EU AI Act written without finding the Digital Omnibus would still describe high-risk obligations applying from 2 August 2026, which they no longer do.

Fabricated sources. The University of Pennsylvania team found 3–13% of URLs fabricated and 5–18% non-resolving across search-augmented models and deep-research agents (arXiv, 3 April 2026). HAQQ's June 2026 benchmark of 3,000 answers found 24% cited or misapplied law that did not support the claim, even though the best model's outright citation fabrication rate was 3%. In Cork v Smith [2026] EWHC 1199 (Ch) the tool invented "Insolvency Rule 12.37(5)" and flagged its own possible inaccuracy; the associate filed regardless.

Agreement with you. Stanford's 2026 AI Index reports that when a false premise is framed as the user's own belief, accuracy collapses (GPT-4o from 98.2% to 64.4% on one benchmark) (Stanford HAI). Ask "confirm that X applies" and the agent will find you reasons.

A prompt that gets the most out of any of them

The safest pattern is two passes: an issue-framing pass with citations forbidden, then, after you have run the real search, an analysis pass restricted to what you verified. The first-pass prompt:

Role: You are a research assistant to a [jurisdiction] [practice area] lawyer.
Task: I am advising on [one-paragraph description of the problem]. Produce a research map, not an answer.
Sources: Use only primary sources and official or court websites you can actually open. Do NOT provide case citations, statutory section numbers or quotations from authorities; instead name the doctrine, the type of authority I should look for, and the search terms I should run in [Westlaw / Lexis / RIS / Beck-online].
Output: (1) the legal issues in order of importance; (2) for each issue, the governing framework at the level of "statute + leading appellate authority", described but not cited; (3) three to five search strings per issue; (4) facts I have not given you that would change the analysis; (5) the jurisdictional limits of everything above.
Constraints: If a point depends on law after your knowledge cutoff, say so. If you are uncertain, say so; do not guess. Plain English, under 900 words.

Then run the search strings yourself, verify what you find, and start the second pass with: "Work only with the authorities I have provided below. Do not add authorities. Analyse ..." followed by the verified extracts. The legal research use case shows the full two-pass workflow with a research log, and the prompt engineering guide explains why forbidding citations in pass one improves the map rather than degrading it.

Grounded alternatives and what "grounded" buys

The legal publishers' answer to all of the above is retrieval from a licensed corpus with a citator check in the loop. As of September 2026: Thomson Reuters' next-generation CoCounsel Legal (US GA 20 August 2026, built on Anthropic's Claude Agent SDK plus TR's own legal model) ships Westlaw Deep Research and a "Deep Research Verify" step against KeyCite; published pricing runs $104–$639 per user per month (Thomson Reuters). Lexis+ with Protégé grounds answers in Shepard's and, since 24 August 2026, runs a "Legal Intelligence Engine" that picks sources and models. Clio Work puts vLex's Vincent (1bn+ documents, 110+ jurisdictions) in front of a solo for $199 a month. A LexisNexis survey published 1 September 2026 found 81% of 543 legal professionals more comfortable with AI grounded in legal sources, up from 70% a year earlier.

Grounding is worth paying for, and it is not a guarantee. Stanford RegLab's 2024 study, the last independent test of this product class, measured hallucination at 17% for Lexis+ AI and 33% for Westlaw AI-Assisted Research (Stanford RegLab); the tools have changed since, and no vendor has submitted to a comparable public test. Thomson Reuters, LexisNexis and vLex all declined to participate in Vals' 2025 research benchmark. Fletcher v. Experian (5th Cir., February 2026) traced 16 fabricated quotations to vLex and CoCounsel, and Lacey v. State Farm (C.D. Cal., May 2025) involved CoCounsel, Westlaw Precision and Gemini. The CoCounsel review, the Protégé review and the Astra for Law page assess each; the RAG glossary entry explains why retrieval narrows the error without closing it.

Verifying a deep-research report before you rely on it

The verification protocol applies to every AI research output; deep-research reports need two extra columns because they cite URLs and they date badly.

CheckWhat you doFail signal
ExistenceOpen each authority in a primary database, not via the report's linkCase name with no reporter hit; section number that does not exist in the consolidated text
SupportRead the passage the report relies onReal case, different proposition (HAQQ's 24%); "misgrounded" citations are the ones the eye misses
TreatmentKeyCite, Shepard's or the national citatorOverruled, distinguished, or amended since the model's cutoff
URLClick every link; note the domainNon-resolving (5–18%) or fabricated (3–13%) URL; a domain that is not the court, regulator or official gazette
JurisdictionConfirm each authority is from the jurisdiction askedUS federal answer to a state question; English authority in an Austrian memo
CurrencyCompare the report's law to today's date, not the cutoffNo mention of a change you know about (a good tell that the agent did not search recent material)
PremiseRe-read your own question for embedded assumptionsReport "confirms" the framing you gave it without testing it
LogRecord tool, model, date, prompt, what you checked, what you rejectedNo log means no defence if a court asks how you verified

Do not delegate any row to the model. A New Jersey lawyer was sanctioned a second time in April 2026 for "using AI to verify AI", and Charlotin's database treats self-validation as an aggravating factor. Citation checkers such as Clearbrief's Cite Check Report (December 2025) can speed up the existence check; they do not do the support check.

A setup that works

For a practice that wants the speed without the sanction risk:

  1. Run deep research only on an enterprise or team tier with training off, and never on a consumer account; on Perplexity personal accounts switch off AI data retention first.
  2. Use the two-pass prompt above: research map with citations forbidden, then analysis restricted to verified authorities.
  3. Keep one grounded tool with a citator in the loop for anything that will be filed or relied on, and open the source anyway.
  4. Add the URL, jurisdiction and currency checks to the firm's verification sheet and log every run.
  5. Re-test the modes on three of your own past questions every quarter; Astra for Law is days old, Gemini 3.5 Pro is still unreleased, and the leaderboard changes monthly. Module 2 of the AI Lab for Lawyers course runs exactly this comparison live, on legal questions the participants bring.

Used this way, a deep-research mode gives a lawyer in 20 minutes what used to take an afternoon of orientation, and the afternoon goes into the authorities instead. The can AI do legal research without hallucinating question has the short version: not yet, and it does not need to, provided you never file what you did not open.

Frequently asked questions

Can I use ChatGPT deep research for legal research?

For orientation, yes; for citations, not without a grounded check. Deep-research modes read many pages and write a sourced report, and on Vals' 2025 US research study ChatGPT Team scored 80% weighted accuracy against a 69% lawyer baseline. But a 2026 University of Pennsylvania study found deep-research agents fabricate 3–13% of the URLs they cite, and courts sanction the filer, not the tool.

Which deep research tool is best for lawyers?

As of September 2026 there is no independent head-to-head on legal questions. Claude Opus 5 and Fable 5.1 lead the Vals Index and rank second and third on its Legal Research Bench; Astra for Law scored 54% on that bench against 38.7% for plain GPT-6 web search, on OpenAI's own numbers; Perplexity's Computer for Counsel routes to 20 or more models. Choose on data policy and reach first, then test on your own matters.

Is Perplexity safe for confidential legal research?

On an Enterprise plan, Perplexity says data is never used for training by default, and Enterprise Pro costs about $40 a seat. On a personal Pro account you should switch off AI data retention in Settings before entering anything about a matter, and even then treat the query as leaving the firm. The same rule applies to consumer ChatGPT and Gemini, which train on conversations unless you opt out.

How do I verify a deep research report before relying on it?

Open every source. For each authority check existence, support and treatment: that the case or provision exists, that it says what the report claims, and that it has not been overruled or amended. Check the URL resolves, the jurisdiction matches the question and the law is current as of today rather than the model's cutoff. Log what you checked. Never ask the model to verify itself.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion