Comparisons
Claude vs Gemini for Legal Work: Context, Workspace and Data Terms
Claude vs Gemini for legal work: close on benchmarks, far apart on data terms. Workspace vs Microsoft fit, Gemini Notebook vs Claude Projects, EU residency.
On this page
Claude and Gemini both read roughly a million tokens and land within a point of each other on LegalBench. The difference that decides the question for a law firm is where the text goes afterwards: consumer Gemini chats are kept 18 months and a subset is read by human reviewers, so Gemini belongs in a Workspace or Enterprise account or not at all. Claude belongs in a Microsoft firm; Gemini in a Google one.
Where they are genuinely equal
Google's flagship as of September 2026 is Gemini 3.1 Pro, released 19 February 2026, with a one-million-token input window and API pricing of $2/$12 per million tokens (rising to $4/$18 above 200,000 tokens). Gemini 3.5 Pro was announced at I/O on 19 May 2026 for "next month" and had still not shipped by late August, while the Flash line reached 3.8 Flash on 2 September. Anthropic's line is Claude Fable 5.1 ($10/$50), Opus 5 ($5/$25) and Sonnet 5 ($2/$10), all with a one-million-token context and 128,000-token output.
On the academic LegalBench (16 September 2026), Fable 5 scores 88.56%, Fable 5.1 88.51% and Gemini 3.1 Pro 87.40%. That is a rounding error in practice. Both models will read a 400-page lease portfolio in one pass, both handle scanned PDFs by vision, and both will summarise a deposition transcript in under a minute. A context window of a million tokens is table stakes at the top of the market now; it stopped being a differentiator in early 2026.
Where they diverge: three tests, one clear pattern
The pattern across the tests that matter for legal work is that Gemini keeps pace on knowledge and falls behind on multi-step execution.
| Benchmark | Claude | Gemini | Source |
|---|---|---|---|
| Vals Index, 58 models (15 Sep 2026) | Fable 5.1 first, Opus 5 second | Not in the top three | vals.ai |
| LegalBench (16 Sep 2026) | Fable 5.1 88.51% | Gemini 3.1 Pro 87.40% | vals.ai |
| GC AI In-House Legal Bench (May 2026, vendor-run) | Opus 4.7 68.4% | Gemini 3.1 Pro 57.5% | gc.ai |
| Harvey LAB all-pass, held out (18 Sep 2026) | Fable 5 13.3%, Opus 5 11.7% | Gemini 3.1 Pro 0.0%; Gemini 3.5 Flash 0.8% (May run) | llm-stats.com |
| Vectara summarisation hallucination (May 2026) | Not in top ranks | Gemini 2.5 Flash Lite 3.3% | github.com/vectara |
The LAB result is the one to dwell on. Harvey's Legal Agent Benchmark asks a model to complete 1,200-plus long-horizon tasks across 24 practice areas and counts a pass only when every one of the expert-written rubric criteria is met. Fable 5 passes 13.3% of tasks that way; Gemini 3.1 Pro passes none. That does not make Gemini useless for legal work, but it does mean you should not give it an unsupervised eight-step drafting workflow yet. Ask it questions; do not hand it a matter. The benchmarks guide explains why the strict metric matters more than criteria-level pass rates in the high eighties.
There is a counter-example worth knowing. A Texas family lawyer rebuilt a child-support calculator he once sold on the App Store in under ten minutes with Gemini, after an attempt with Claude failed. On a single, well-specified build, either model can win on the day.
Workspace or Microsoft: the ecosystem question usually settles it
Most firms do not choose a model. They choose a productivity suite, and the model follows.
A Google Workspace firm gets Gemini inside Gmail, Docs, Drive and Meet through a Workspace add-on, plus Gemini Enterprise for agents: Skills went GA on 17 June 2026, an Agent Registry with egress allow and deny lists arrived on 25 June, and the Workflow Builder went GA on 3 September 2026, according to Google Cloud's release notes. Google also demonstrated a "Gemini Enterprise for Legal" configuration on 17 September 2026. Note that the consumer Google AI Pro and Ultra subscriptions are personal-account only; Workspace customers need the add-ons.
A Microsoft firm gets Claude in more places than it may realise. Claude runs in Microsoft 365, in a first-party Word add-in with tracked changes, in Claude for Chrome and in Cowork, which merged into the main Claude app on 17 September 2026 alongside Claude Docs and Slides in beta. Microsoft's own Word Legal Agent (30 April 2026) requires Anthropic enabled as a subprocessor, and Copilot Cowork (GA June 2026) integrates Anthropic models. Freshfields' CIO Gil Perez put the industry's position bluntly: legal "lacks optionality. It's a Microsoft dominated field." Freshfields then signed with Google Cloud and, a week later on 23 April 2026, with Anthropic, so even the most Google-committed firm hedged.
If you are a Google firm, start with Gemini and add Claude via the API or Vertex AI for the drafting tasks where it is stronger. If you are a Microsoft firm, start with Claude and use Gemini Notebook for closed-universe file work. The Gemini review and Claude review cover the individual products.
Gemini Notebook vs Claude Projects
These are the two features lawyers ask about most, and they solve different problems.
| Gemini Notebook (ex-NotebookLM) | Claude Projects plus Skills | |
|---|---|---|
| Renamed / launched | 16 Jul 2026; 30M+ users; 600,000+ organisations | Projects since 2024; Skills across Claude.ai, Claude Code and the API |
| Core idea | Closed universe: answers only from uploaded sources, cites passages | Persistent context plus instructions in one workspace; Skills fire on trigger phrases |
| Source limits | 50 sources per notebook free, 300 on paid tiers | Governed by the 1M context and project storage |
| Best legal use | Case file: one-page issues brief tied to passages, chronology, cross-examination outlines | Matter workspace: playbook, precedents, style guide, then repeatable skills such as a redline or a memo format |
| Open-web research | No; it is deliberately closed | Yes, with web search and connectors |
| Code execution | Each notebook has a "secure cloud computer" | Via Cowork, Claude Code or the analysis tool |
| Enterprise residency | EU residency and ML processing for NotebookLM Enterprise since 3 Jun 2026 | Vertex AI and Bedrock regions; Compliance API for audit |
| Watch out for | Spot-check every citation; use only in Workspace or enterprise accounts | Skills are single-user markdown unless the firm curates them; no built-in citation check without a research connector |
The Gemini Notebook guide walks through a litigation file setup, and the Claude Projects playbook does the same for a transaction. A useful rule: Notebook for reading, Projects for producing.
Data terms: the section to show your risk partner
Everything above is secondary to this. Google's Gemini Apps privacy page (updated 10 August 2026) states that consumer Gemini activity is kept 18 months by default, that a subset of chats is read by human reviewers, that chats are used to train models unless "Keep Activity" is turned off, that reviewed chats are retained up to three years, and that even with activity off, chats are kept for 72 hours. It then says:
"Please don't enter confidential information that you wouldn't want a reviewer to see." — Google, Gemini Apps Privacy Hub, support.google.com
Workspace and Vertex AI accounts are excluded from training by default through the admin console. That is the whole reason a law firm must never let a lawyer use a personal Gemini account for client material. The UK Upper Tribunal in UK v SSHD [2026] UKUT 81 (IAC) said pasting client letters into public AI tools puts them "in the public domain"; the SRA's Warning Notice of 17 August 2026 cites that decision.
Claude's terms are stricter at the consumer level and cleaner at the business level. Since 28 August 2025, Free, Pro and Max users must choose whether chats train models; opted-in data is kept five years, opted-out data 30 days, and there is no human-review programme of the Gemini kind described in the briefs. Team, Enterprise, API, Bedrock and Vertex content is never used for training. Two caveats from Anthropic's retention documentation: Fable 5.1 and Mythos 5.1 require 30-day retention and are not zero-data-retention eligible without express authorisation, and the Team and Enterprise chat interfaces are not ZDR-eligible either.
| Question a DPO will ask | Consumer Gemini | Workspace / Gemini Enterprise | Claude consumer | Claude Team / Enterprise / API |
|---|---|---|---|---|
| Used for training? | Yes unless Keep Activity off | No by default | Your choice; 5-year retention if yes | Never |
| Human review of chats? | Yes, a subset | Governed by contract | Not described | Not described |
| Default retention | 18 months; 72 hours minimum | Admin-controlled | 30 days if opted out | Not retained by default on API; Fable 5.1 30 days |
| EU residency | No | Yes for enterprise tiers | No | Vertex / Bedrock regions |
| Fit for client data | No | Yes with DPA | No | Yes with DPA and policy |
For the GDPR side, including the Article 28 processor terms and the DPIA question, see GDPR and AI tools for lawyers. The step-by-step privacy settings page shows where the Keep Activity and model-improvement toggles live.
Recommendation by scenario
| Scenario | Use | Notes |
|---|---|---|
| Google Workspace firm, general drafting and email | Gemini via Workspace add-on | Add Claude on Vertex AI for redlines and memos |
| Microsoft 365 firm | Claude Team or Enterprise | Word add-in, Cowork, Compliance API; Gemini Notebook for file reading if permitted |
| Litigation file of 200 documents, closed universe | Gemini Notebook (Enterprise) | Chronology and issues brief; verify every passage cite |
| Repeatable matter workflow (playbook, precedents, memo format) | Claude Projects with Skills | The Projects playbook has the template |
| Unsupervised multi-step agent | Claude, with approval gates | Gemini 3.1 Pro scored 0.0% on LAB all-pass |
| Deep research with citations | Either, then open every source | Search-augmented agents fabricated 3–13% of URLs in a University of Pennsylvania study (April 2026); see deep research tools compared |
| EU firm needing regional processing | Gemini Enterprise or Claude on Vertex AI in an EU region | Consumer tiers of both are out |
| Any lawyer with only a personal Gemini account | Stop using it for client work today | Google's own privacy page says so |
A setup that works
For a Workspace firm of twenty lawyers: Gemini Workspace add-on for every seat; Gemini Enterprise with Workflow Builder for two or three defined agents; NotebookLM Enterprise in an EU region for case files; and Claude Opus 5 through Vertex AI, wired into a Projects-style prompt library, for contract redlines and memos. Write the policy first, then buy the seats.
For a Microsoft firm of the same size: Claude Team on the premium seat for drafters and standard for everyone else; Claude for Legal plugins with the iManage or NetDocuments connector; Copilot for email; and a single Gemini Enterprise workspace only if a practice group insists on Notebook. Module 3 of the AI Lab for Lawyers course covers the file-handling and multimodal differences between the two products on real documents.
Whichever route you take, the citation duty is identical. Lacey v. State Farm (C.D. Cal., 6 May 2025) cost Ellis George and K&L Gates $31,100 for nine bad citations produced with CoCounsel, Westlaw Precision and Google Gemini. The verification protocol applies to Gemini, Claude and everything else.
Frequently asked questions
Is Gemini or Claude better for lawyers?
For drafting and multi-step legal work Claude is ahead as of September 2026: it leads the Vals Index and scores 13.3% on Harvey's strict agent benchmark where Gemini 3.1 Pro scores zero. Gemini is the better fit for firms living in Google Workspace and for closed-universe case-file work in Gemini Notebook. On the academic LegalBench the two are within about one point.
Can lawyers use Gemini for confidential client documents?
Only through a Workspace, Gemini Enterprise or Vertex AI account where training is excluded by the admin console. The consumer Gemini app keeps activity for 18 months by default, human reviewers read a subset of chats, reviewed chats are kept up to three years, and Google's own privacy page says not to enter confidential information you would not want a reviewer to see.
What is the difference between Gemini Notebook and Claude Projects?
Gemini Notebook, renamed from NotebookLM in July 2026, is a closed-universe tool: it answers only from the sources you upload, up to 300 per notebook on paid tiers, and now has a secure cloud computer for code. Claude Projects hold static uploaded context and custom instructions in a workspace and combine with Skills, which fire on trigger phrases across Claude.ai, Claude Code and the API.
Does Gemini offer EU data residency for law firms?
Yes for enterprise products. Google added EU data residency and in-region ML processing for NotebookLM Enterprise on 3 June 2026, and Gemini Enterprise runs on Google Cloud regions. Anthropic offers Claude through Vertex AI and Bedrock regions and per-organisation zero-data-retention on the API, though Fable 5.1 requires 30-day retention. Check the DPA, not the marketing page.
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.
Related reading
Tools
Gemini for Lawyers: Workspace Integration, Gemini Notebook and Data Caveats
Gemini for lawyers as of September 2026. Why Workspace is the only safe door, what Gemini 3.1 Pro and Gemini Notebook do well, and where Google's terms bite.
7 min read
Tools
Gemini Notebook (NotebookLM) for Lawyers: A Grounded Case-File Assistant
Gemini Notebook (ex-NotebookLM) for lawyers as of September 2026. How source-grounded answers work on a case file, three workflows, source caps and data terms.
6 min read
Playbooks
Set Up Claude Projects for a Matter: Instructions, Files and Memory
Set up a Claude Project for a legal matter in 30 minutes: sample instruction text, which files to upload, Skills, memory, Team sharing and three prompt flows.
10 min read
Tools
Claude for Lawyers: An Honest Review of Anthropic's Models for Legal Work
Claude for lawyers, reviewed as of September 2026: Fable 5.1, Opus 5 and Sonnet 5, plan prices, data retention by plan, legal plugins, benchmarks, weaknesses.
9 min read
Regulation & Ethics
GDPR and AI Tools for Lawyers: Lawful Basis, DPIAs and Vendor Terms
How GDPR applies when law firms use Claude, ChatGPT or Copilot: controller and processor roles, lawful basis, DPIA triggers, transfers and vendor terms.
11 min read
Glossary
Context Window — Definition for Lawyers
A context window is the text an AI model can hold at once. As of September 2026 it is 1 million tokens on Claude, GPT-6 and Gemini, about 555,000 words.
4 min read