AI Agents for Lawyers

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.

By Dr. Niklas SchmidtUpdated 13 November 20269 min read
On this page
  1. What you get in September 2026
  2. Plans and what they cost
  3. Who can see your prompts: the data policy by plan
  4. The features that matter for legal work
  5. How Claude scores
  6. Where Claude falls short
  7. A firm setup that works

Claude is, as of September 2026, the strongest general-purpose model family for legal work on most public benchmarks and the one with the deepest legal tooling: 12 open-source practice plugins, more than 20 document-system connectors and a desktop agent. It is also a family whose best model cannot run under zero data retention, and whose consumer plans train on your chats unless you opt out. Both facts should shape how a firm uses it.

What you get in September 2026

Anthropic's current line-up, per its model overview, is four models. Three of them share a 1-million-token context window, which Anthropic puts at roughly 555,000 words, or several data rooms' worth of leases in a single conversation.

ModelReleasedAPI price per million tokens (in / out)Context / max outputWhere you meet it
Claude Fable 5.11 Sep 2026$10 / $501M / 128KPaid plans only; not on Free
Claude Opus 524 Jul 2026$5 / $251M / 128KDefault on Max, strongest on Pro
Claude Sonnet 530 Jun 2026$2 / $101M / 128KDefault on Free and Pro
Claude Haiku 4.52025$1 / $5200K / 64KAPI, bulk classification

Two things about the models themselves matter to a lawyer. First, thinking is now "adaptive" and set with an effort dial rather than a separate reasoning model, and on Fable 5.1 the raw chain of thought is never returned, so you will not see the model's working. Second, Opus 5 changed its habits: Anthropic's release notes say it writes longer deliverables, delegates to sub-agents more readily and "verifies its own work without being told to". Prompts written in 2025 that end with "now add a final verification step" now cause double-checking loops. Anthropic's Opus 5 announcement claims first-turn NDA redlines "nearly double Opus 4.8" and the largest gains "in practice areas like corporate governance and arbitration"; those are vendor numbers, but they match what the independent redlining benchmarks show below.

There is also a lesson in dependency. On 12 June 2026 the US government restricted Fable 5 access for foreign nationals, and Anthropic suspended access for everyone for 18 days because it "had no reliable way to verify nationality in real-time". If a live matter depends on one model, keep a second one warm.

Plans and what they cost

PlanPriceWho it suits
Free$0Trying Sonnet 5; never client data
Pro$20/month ($17 annual)A solo who has turned training off
Max$100 or $200/monthHeavy users who want Opus 5 by default and Cowork
Team$25 standard / $125 premium per seat, 5-seat minimumMost small and mid-sized firms
EnterpriseCustomFirms that need SSO, audit and ZDR on Claude Code
APIMetered per tokenFirms building their own tools or using vendor platforms

The premium Team seat is what gives a lawyer enough Opus 5 and Fable 5.1 usage for a day of document work; the standard seat runs out quickly on long contracts.

Who can see your prompts: the data policy by plan

This is the section to read twice. The rules come from Anthropic's consumer terms update and its API retention documentation.

SurfaceTrains on your content?RetentionZero data retention available?
Free, Pro, MaxOnly if you opt in (you must choose since 28 Aug 2025)5 years if opted in; 30 days if opted outNo
Team, Enterprise chatNeverStandard retention; flagged content up to 2 yearsNo
Claude Code under EnterpriseNeverConfigurableYes
API, Bedrock, VertexNeverNot retained by defaultYes, arranged per organisation, except Covered Models
Fable 5.1 and Mythos 5.1 ("Covered Models")Never30-day retention requiredNot without express authorisation
Compliance API feed, local Cowork and Claude Code transcriptsNever6 years by defaultn/a

The uncomfortable consequence: the strongest Claude model is the least private one. If your engagement terms or your bar's guidance require zero data retention, your ceiling is Opus 5, not Fable 5.1. Team and Enterprise chat interfaces are not ZDR-eligible at all. And the six-year default for Compliance API transcripts is a discovery consideration in itself. The confidentiality checklist sets these numbers beside the other vendors.

One court has already tested the consumer tier. In United States v. Heppner (S.D.N.Y., Judge Rakoff, opinion 17 February 2026) a defendant's research memos drafted in consumer Claude were held not privileged and not work product: "Claude is not an attorney", and Anthropic's terms allow logging and disclosure. Rakoff signalled that privilege might attach if counsel directs the AI use. The full analysis is in privilege and AI.

Projects hold static context (a matter summary, the client's style guide, the governing law) that every chat in the Project sees. Set-up is covered in Claude Projects for a matter.

Skills are folders with a SKILL.md file plus optional scripts and reference documents; they fire only when a prompt matches their trigger and work across Claude.ai, Claude Code and the API. A US small-firm lawyer's post about hand-rolled Skills that edit Word XML for tracked changes drew about 7 million views in March 2026, according to Artificial Lawyer.

Cowork is the desktop agent with sandboxed file and shell access. It launched as a research preview in January 2026 and was merged into the main Claude app on 17 September 2026, alongside Claude Docs and Claude Slides in beta. Anthropic reported legal professionals as the most engaged Cowork users of any function. It also produced the year's most quoted mishap: a venture investor let Cowork "organise" a desktop, approved a deletion, and lost 15 years of family photos to an rm -rf until an iCloud rollback recovered them. The Cowork review explains what to let it touch.

Claude for Legal began as one plugin on 3 February 2026, the day Thomson Reuters fell 18 percent intraday and RELX had its steepest one-day drop since 1988. On 12 May 2026 it became 12 practice-area plugins with 90-plus agents and skills and more than 20 MCP connectors (iManage, NetDocuments, DocuSign, Ironclad, Relativity, Everlaw, Thomson Reuters CoCounsel, LexisNexis, Harvey and others), all open source under Apache 2.0 at github.com/anthropics/claude-for-legal. Every plugin begins with a "cold-start interview" that writes your practice profile, tags unverified citations "[verify]" and gates sending or filing behind a human confirmation.

MCP connectors are how Claude reaches your documents. The iManage MCP Server (14 May 2026) runs permission-bound and logged and "respects existing ethical walls"; NetDocuments says "documents never leave the platform". See connecting Claude to your DMS.

Claude for Chrome went GA on 18 December 2025 for Pro, Team and Enterprise, with admin allowlists. Anthropic's mitigations cut prompt-injection success from 23.6 to 11.2 percent across 123 scenarios; Zenity's "PleaseFix" research still showed a zero-click hijack through a poisoned email. Read agent security for lawyers before switching it on.

How Claude scores

BenchmarkDateClaude resultContext
Vals Index (58 models)15 Sep 20261. Fable 5.1, 2. Opus 5GPT-6 Astra third
Vals Legal Research Bench15 Sep 2026Opus 5 second, Fable 5.1 thirdMeta's Muse Spark 1.3 Max first
Academic LegalBench16 Sep 2026Fable 5 88.56%, Fable 5.1 88.51%Gemini 3.1 Pro 87.40%
Harvey BigLaw BenchMay 2026Opus 4.7 90.9%GPT-5.6 Sol 92.7% (Jul 2026)
Harvey LAB, strict all-criteria pass18 Sep 2026Fable 5 13.3%, Opus 5 11.7%GPT-5.5 2.1%, Gemini 3.1 Pro 0.0%
HAQQ (3,000 answers, 51 practice areas)5 Jun 2026Opus 4.8 first, 30.02/35Citation hallucination about 12% vs GPT-5.5 at 3%
GC AI In-House Legal Bench (vendor-run)May 2026Opus 4.7 68.4%GC AI 86.8%, GPT-5.5 79.8%

Two readings follow. Claude is the model to beat on legal reasoning and on long-horizon agent tasks, where Harvey's own LAB leaderboard puts Fable 5 first. It is not the lowest hallucinator: GPT-5.5 fabricates fewer citations, at the cost of being the slowest and dearest model in the HAQQ test. The benchmarks guide explains why a 13.3 percent "all-pass" score and a 90.9 percent "BigLaw Bench" score describe the same model.

Where Claude falls short

  1. Retention on the flagship. Covered Models require 30-day retention; a firm that has promised ZDR must use Opus 5.
  2. No ZDR on Team or Enterprise chat. Only Claude Code under Enterprise and the API qualify.
  3. Refusals. Fable returns a "refusal" stop reason when a classifier declines. During the July 2026 OpenAI sandbox-escape incident, Hugging Face's responders could not get Claude to help reverse-engineer attacker payloads and fell back to an open-weight model. For criminal-defence or cyber-incident work, expect the occasional wall.
  4. It still fabricates. On 15 May 2025 Anthropic's own outside counsel, Latham & Watkins, apologised in Concord v. Anthropic after Claude was asked to format a citation and kept the link but invented the author and title; the judge distinguished "a missed citation and hallucination generated by AI". In Damien Charlotin's database of court decisions, where the tool is named, Claude accounts for about 15 percent. The verification protocol is not optional.
  5. Gaps a vertical tool fills. GC AI's critique of the legal plugin names three: context that resets per session, no citation verification without a connector, and single-user markdown instead of team playbooks.
  6. Behaviour drift. Opus 5 verifies unprompted and narrates progress; every model release changes how your saved prompts behave.

A firm setup that works

  1. Choose the surface by data need: Team with premium seats for most firms; Enterprise if you need Claude Code under ZDR; API through a vendor platform if you need ZDR at scale. Never client data on a consumer plan.
  2. On any consumer account that exists in the firm, turn training off under Settings, Privacy, before anything else.
  3. Create one Project per matter. Paste instructions like the following:
You are assisting a team of Austrian and English-qualified lawyers on the matter
described in the attached summary. Governing law for the SPA is English law.
Work only from the documents in this Project and from authorities the team
pastes in; do not supply case citations from memory. When a point needs
authority, write [AUTHORITY NEEDED] and describe what would be needed.
Output: British English, numbered paragraphs, defined terms as in the SPA.
Flag any assumption you make in a separate list at the end.
  1. Install the relevant Claude for Legal plugin and run its cold-start interview with a partner in the room; the profile it writes is the playbook every later skill reads.
  2. Connect the DMS through MCP with the admin toggle, not by uploading documents by hand.
  3. Keep the research log and the citation check outside Claude.
  4. Record who was trained on what: the EU AI Act's Article 4 duty expects an internal training record, and the state of AI in legal practice shows how few firms have one.

Module 2 of the AI Lab for Lawyers course builds this Project-and-plugin setup live, from the first instruction file to a verified memo. Freshfields, which rolled Claude out to 5,700 staff across 33 offices in April 2026 and saw usage grow about 500 percent in six weeks, reached the same conclusion by a longer route: the model is good, the setup decides whether it is safe. How it compares with the alternative is in Claude vs ChatGPT for lawyers.

Frequently asked questions

Is Claude safe to use for confidential client work?

On Team, Enterprise or the API, yes, subject to your own risk assessment: Anthropic never trains on that content and does not retain API content by default. On Free, Pro or Max you must switch training off, and your chats are still retained for 30 days. Consumer-plan research has been held not privileged in United States v. Heppner (S.D.N.Y., February 2026).

Which Claude model should a lawyer use?

As of September 2026, Claude Opus 5 is the practical default: it is the strongest model available under zero data retention, it is the default on Max, and Anthropic reports its first-turn NDA redlines nearly doubled Opus 4.8. Fable 5.1 scores marginally higher but requires 30-day retention. Sonnet 5 is enough for summaries and first drafts.

Does Claude hallucinate case law?

Yes. In the HAQQ benchmark (June 2026) Claude Opus 4.8 ranked first overall but still produced citation hallucinations in roughly 12 percent of answers, and Anthropic's own counsel had to apologise in May 2025 for a Claude-corrupted citation. Use it for analysis and drafting, verify every authority in Westlaw, Lexis or the court record, and keep a research log.

How much does Claude cost for a law firm?

Pro is $20 a month ($17 annual), Max $100 or $200, and Team $25 per standard seat or $125 per premium seat with a five-seat minimum. Enterprise is custom. API usage is metered: Opus 5 costs $5 per million input tokens and $25 per million output tokens, Sonnet 5 $2 and $10, Fable 5.1 $10 and $50.

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