Claude for Legal Artifacts: What the Gallery Does and What It Cannot Do

Last Updated: 29 October 2026

Anthropic opened a gallery of ready-made legal artifacts inside Claude for Legal on 28 October 2026. The announcement did 3.99 million views in under a day, which tells you the market was waiting for something. The gallery itself is more interesting than the launch post admits, and less capable than the replies assume: it is a first-pass machine with sharing guardrails, sitting on top of a benchmark that frontier models still fail more than nine times out of ten.

What is the Claude for Legal artifacts gallery?

The gallery is a set of 19 pre-built artifacts inside Claude for Legal: interactive trackers, boards and tables that open with a worked example, take your own documents and connected sources, produce a first pass, and link every output back to its source. Six are built by Anthropic, four by legal tech companies and law firms, and nine support access-to-justice work.

Anthropic says the two most common requests from legal teams since the May launch were simpler setup and easier ways to work with colleagues. The gallery answers both at once: instead of prompting from a blank page, you start from a working example and adapt it.

The six Anthropic-built artifacts cover the work legal teams raised most often:

Artifact	What it does
Cite check	Compares a legal document with its cited sources and flags quotes that do not match, wrong page numbers, and sources that do not support the claim.
Jurisdiction survey matrix	Compares a topic such as data breach notice rules across states or countries in one table, with links to the underlying law.
Matter board	Turns engagement letters, emails or spreadsheets into one board of open matters with owner, next step and next deadline.
Playbook deviation board	Compares contracts with your standard terms, flags each differing clause, and routes it to whoever approves exceptions.
Regulatory horizon scan	Maps upcoming regulations on a timeline for your business and jurisdictions, ranks them by impact, and links the sources.
Tabular review	Builds a table from a set of contracts showing the terms you ask about, such as severance or non-competes, with links to the exact language.

The partner shelf is where the gallery starts to look like a market. Goodwin built an AI Governance Benchmark, Harvey built Harvey Intake (business users send requests in Claude, routine ones are answered from Legal's own guidance, the rest route to a lawyer), Foley & Lardner built an Obligations Checker, and PwC built a Subpoena Response Tool. Anthropic says artifacts from Legora and Orrick are coming.

The nine access-to-justice artifacts are the most concrete part of the release. Five were built by nonprofits supported with Claude credits: Zealous (Criminal Charge Reality Check, New York only), Mobile Pathways (Immigration Papers Reader), Rising for Justice (Rent Ledger Reviewer and Record-Sealing Screener, both D.C. law), and The Access Project (Record Clearance). That is a real deployment of the same primitive to people who cannot afford a lawyer, and it is the part of this launch worth watching over the next year.

How does a legal artifact actually work?

An artifact is a small interactive tool that lives inside Claude. You open an example, attach documents or connectors, ask for a first pass, check the source links, and then share it with colleagues, clients or partners under sharing settings the organisation's admin controls.

The workflow Anthropic describes is deliberately narrow, and that narrowness is the point:

1. Start from the example. Every artifact opens with a worked example so you can see how the output comes together before your data goes anywhere near it.
2. Bring your own material. Upload documents or connect a system of record. Anthropic has added more than 60 legal connectors since May, naming Bloomberg Law, Filevine, Ivo, Legora and Patlytics, with iManage, NetDocuments, Docusign, Ironclad and Thomson Reuters described on its legal solutions page.
3. Claude completes the first pass. It produces the draft table, board or scan, with a link to the source behind each answer.
4. A lawyer applies judgment. The output is a reviewable work product, not a filed document. Someone accountable reads it, corrects it and signs off.
5. Share it with guardrails. Colleagues across the business, or clients and partners, can work on the artifact under settings Anthropic says support privacy and ethical walls between teams.

The admin controls are where most firms will spend their first week, and the details matter more than the marketing. External sharing, which lets anyone with the link open an artifact, is a separate setting from email invitations to named people outside the organisation; turning one on does not turn on the other. Turning external sharing off also stops existing public links from working, except for artifacts an admin has individually allowed. Artifacts that use connected apps, or that ask Claude questions, cannot use "anyone with the link" at all.

Two more details worth putting in front of a partner. Usage draws from each user's seat limits, including session and weekly limits, and admins can buy credits for people who need more. And uploaded assets are stored persistently under the organisation's retention policy: an Enterprise plan can set a custom retention period after which an artifact nobody has opened can be permanently deleted. That is a data governance conversation, not an IT one.

Why are legal teams paying attention now?

Because adoption has outrun strategy. Thomson Reuters found 91% of professional services organisations say they are falling short of AI's potential value while only 22% of law firms have a visible AI strategy, and Anthropic's own customers said the blocker was setup and collaboration rather than model quality.

The adoption numbers are no longer the story. Thomson Reuters' 2026 research, surveying 1,816 professionals across more than 60 countries, puts generative AI use at 41% of law firms and 47% of corporate legal departments, up from 28% and 23% a year earlier. Document review, legal research and summarisation each sit around 74% to 77% of use. Wolters Kluwer's companion survey found 92% of lawyers using at least one AI tool daily.

What lags is the operating model. That gap is why the artifact layer matters: it moves the unit of work from a prompt in a chat window to a shared instrument that a team can own, version and audit. It is also why the vendors are racing. Harvey raised US$550M at a US$15.5B valuation in September 2026 on roughly US$400M of annual recurring revenue. Legora raised US$550M at an US$8.8B valuation on 21 October 2026, five months after a US$550M round at US$5.6B. Anthropic is not entering a quiet category; it is entering the most contested one in legal technology.

What are the three questions that decide whether you can share one?

Privilege, verification and confidentiality. Each has a documented answer in 2026 case law or court guidance, and each is decided before anyone clicks share, not after.

1. Privilege: whose circle is the model inside?

In United States v. Heppner, Judge Rakoff of the Southern District of New York held on 17 February 2026 that a defendant's exchanges with the consumer version of Claude, used for legal research, were not protected by attorney-client privilege or the work-product doctrine. Commentary since then has overread the decision as an automatic waiver for any AI use. The narrower and more useful reading is that the tool sat outside the privileged circle: it was a consumer service, not a channel to counsel.

The practical consequence for a gallery artifact is straightforward. If the artifact runs in an organisation's own Claude workspace, under its retention settings, with audit trails and access control, the analysis is materially different from a lawyer pasting a client file into a personal account. That is not a guarantee of privilege. It is the difference between a defensible position and no position at all.

2. Verification: who signs the document?

Australia's Federal Court moved first and moved hard. Practice Note GPN-AI, which commenced on 16 April 2026, requires that where generative AI has been used in preparing a filed document, the responsible person confirms the accuracy of the content and the citations, and that the document discloses its use where required. Clause 5.1 warns that non-compliance can attract adverse costs orders.

An artifact helps here, but only partly. Source links make a cite check faster; they do not discharge the responsible person's duty. The Cite check artifact exists precisely because quotes, page numbers and propositions still go wrong. Treat the artifact as a first pass that makes human verification cheaper, and keep the confirmation where the rules put it: on a named person.

3. Confidentiality: is the sharing setting a policy?

GPN-AI clauses 4.13 to 4.15 tell users to consider confidentiality, suppression and privacy restrictions before information goes into a generative AI tool. Footnote 7 notes that the risk of inadvertent disclosure may be lower in closed or controlled environments, while making clear the Court still expects rigour. In Australia, APP 8 adds a cross-border layer: before disclosing personal information to an overseas recipient, an entity must take reasonable steps to ensure the APPs are not breached, and remains accountable for what happens next.

This is where the ethical walls claim deserves scrutiny. A sharing setting can support a wall; it is not a wall. Matter-level permissions, conflict checks, access logs and a documented policy on what may leave the firm are the actual controls. An artifact that carries client confidences is a disclosure decision every time it is shared externally, including when the recipient is the client.

What do the benchmarks actually say about that first pass?

On Harvey's Legal Agent Benchmark holdout, frontier models complete less than 10% of matters end to end under an all-pass standard. The best result is Claude Opus 4.7 at 7.1%, at about US$50.90 and 22 minutes per task. Narrower benchmarks look far better, which is exactly the point: the number depends on the task universe you measure.

LAB is deliberately strict. Each task starts from a partner-style instruction inside a closed-universe client matter, and a task passes only if every rubric criterion passes. Under that standard, Opus 4.7 leads at 7.1%, Sonnet 4.6 sits at 5.4%, Opus 4.6 at 4.2%, GPT-5.5 at 2.1% and Gemini 3.5 Flash at 0.8%. Harvey also found that no single model leads every practice area, and that the strongest configurations are multi-model by design.

Narrower benchmarks tell a different and equally true story. Claude Opus 5.5, released 22 September 2026, scores 3.75% on LAB (ranked 31st of 64 in Vals AI's tracking) but 50.48% on Vals' Legal Research Bench, fourth of 68 models, and 48.4% task pass on Legal Benchmarks' independent contract workflow set, at 91.7% criteria pass, 93.9% form and US$1.47 per task in about three minutes. Research and contract review are close to solved as first-pass tasks. End-to-end matter handling is not.

The most useful finding in Harvey's work is behavioural rather than numerical. Its trace analysis found that agents which revise after checking score 1.5 points higher on all-pass, agents that run a validation step after drafting gain 0.8, and agents that draft without review lose 1.2. High parallel tool use, five or more calls in a turn, costs 0.5. In other words, the gains come from the loop, not the model: read the matter, research what is missing, draft, check against the instruction, revise.

That is the strongest argument for artifacts as a product category. An artifact can encode the loop. A chat window cannot. If you are building or buying one, the question to ask is not which model it uses; it is which review step it forces.

What did Anthropic leave out of the announcement?

The connector directory is a commercial map. In May 2026, pull request 5 in Anthropic's public claude-for-legal repository removed every Lexis reference at LexisNexis's request: 97 files changed, Lexis+ Protégé entries stripped from ten plugins, and the [Lexis+] source tag replaced with [CourtListener] throughout. Thomson Reuters, Westlaw and CoCounsel were left untouched because Thomson Reuters is an active partner.

None of that is improper. Partners get to decide whether they appear in a vendor's directory, and a company can ask to be removed from a competitor's showcase. But it does mean the gallery is not a neutral map of the legal data market, and the repository is not a community project in the ordinary sense. It is Apache-2.0 licensed, has 9,822 stars and 1,921 forks, and 566 files, and its contents move when partners say so.

The practical lesson for a firm choosing an artifact stack is to check one thing before anything else: is your system of record a first-class connector, or a workaround? If your matters live in a system the artifact cannot reach, the artifact will quietly become another place documents get copied to, which is precisely the sprawl it was meant to end.

What would a local-first version of this look like?

The same pattern, with different custody. Post-trained open-weight models are now landing inside the closed-source frontier band on legal work, and Harvey's own research makes the governance case for hosting them inside a firm's environment rather than a vendor's.

Two results matter. Harvey and Baseten Research post-trained a 27B open-weight model and brought it into the closed-source frontier band on LAB, using iterative supervised fine-tuning inside a compaction harness. Separately, Harvey and Applied Compute post-trained GLM-5.1 and lifted its rubric pass rate from 0.853 to 0.913 and its all-pass rate from 0.059 to 0.126, beating GPT-5.5 xhigh and Opus 4.8 Max on rubric pass rate and approaching Opus 4.8 Max on all-pass. Harvey's own summary of why this matters is worth quoting in full: open-weight models can be hosted within a firm's own secure cloud environment and enable deeper auditability through full visibility into the model's reasoning traces, tool calls and intermediate decisions.

That is the same argument we make to clients, from the other direction. On a pair of DGX Sparks we run DeepSeek V4.1 Flash at about 35 tokens per second on a single stream and about 62 tokens per second at four concurrent streams, with the matter data never leaving the building. It is not a frontier model on the hardest matters. It is entirely adequate for the first-pass work that makes up most of an artifact's value: extraction, tabulation, comparison, summarisation and cite checking.

The honest cost framing matters here, because it is where most on-premise pitches fall apart. Two Sparks cost more than a year of Claude Team seats for a small team, so the case cannot be licence savings. It rests on four things: work that cannot leave the building, permitted-work guarantees you can evidence, API volume high enough that per-token pricing dominates, and a shared staff of dozens where one local inference layer serves everyone. If none of those apply, the subscription is the better buy.

What is a decision rule you can use this week?

Classify each task as cloud-safe, hybrid or never-cloud, then let the classification pick the tool. Most firms get this backwards: they pick the tool first and negotiate the data question afterwards.

• Cloud-safe: public law, published judgments, firm templates, marketing copy, training material. Use the gallery artifacts freely.
• Hybrid: client matter documents where the client has consented to AI processing, or where the tool is under an enterprise agreement with retention and audit. Use artifacts with external sharing off, and keep the matter in the system of record.
• Never-cloud: suppressed or non-publication material, privileged advice, unredacted personal information, matters under a confidentiality order, and anything a client has expressly prohibited. Local inference or nothing.

Then run a five-step pilot, on one matter, in one week:

1. List the ten tasks your team repeats most, and classify each one.
2. Pick one cloud-safe task and adapt the matching gallery artifact to your house style.
3. Run it on a closed matter file and check every source link by hand. Count the errors.
4. Write down what the lawyer had to fix. That list is your evaluation set.
5. Only then decide on sharing: who may invite externally, what the retention period is, and who reviews the log.

The firms that get value from this release will not be the ones that switch it on fastest. They will be the ones that can show, six months later, which artifacts touched which matters, who checked the output, and why the data was allowed to travel where it did.

Frequently asked questions

What is Claude for Legal?

Claude for Legal is Anthropic's package for legal teams, launched in May 2026. It combines Claude with legal-specific connectors, practice-area plugins and, since 28 October 2026, a gallery of ready-made artifacts. It runs on Team and Enterprise plans, with admin controls for sharing, retention and usage.

Are the legal artifacts included in my Claude plan?

The gallery is part of Claude for Legal on Team and Enterprise plans. Artifact usage draws from each user's seat limits, including session and weekly limits, and admins can buy additional credits. There is no separate artifact licence.

Can I share a legal artifact with a client?

Technically yes, if external sharing is enabled or the client is invited by email. Whether you should is a confidentiality decision: GPN-AI clauses 4.13 to 4.15 require you to consider restrictions before information enters a tool, and in Australia APP 8 governs cross-border disclosure of personal information. Treat every external share as a disclosure, not a convenience.

Does Claude train on my legal documents?

On Team and Enterprise plans, Anthropic states that customer data is not used for model training by default, and uploaded assets fall under the organisation's retention and deletion policy. Consumer accounts are a different product with different terms, which is exactly the distinction that mattered in Heppner.

Can artifacts replace a lawyer?

No. On Harvey's Legal Agent Benchmark holdout, the best frontier model passes 7.1% of matters end to end. Artifacts make the first pass faster and more consistent; they do not remove the review, the judgment or the signature.

What is the Legal Agent Benchmark?

LAB is Harvey's open-source benchmark for legal agents. Tasks begin with a partner-style instruction inside a closed-universe client matter, and a task passes only if every criterion in an expert-written rubric passes. That all-pass standard is why LAB scores look low next to research or contract-review benchmarks.

What to watch next

Three things will decide whether this gallery becomes infrastructure or a demo shelf. The first is whether Legora and Orrick artifacts ship with the same source-link discipline. The second is whether LAB's leaderboard, which Harvey is building with Artificial Analysis, starts to separate harness quality from model quality, because that is the number buyers actually need. The third is whether firms publish their own artifacts, which is the moment the gallery stops being Anthropic's product and becomes the profession's.

If you want to talk through which of your tasks are cloud-safe, hybrid or never-cloud, and what a local-first version would cost to run, that is the work we do at Flowtivity. Bring one matter file and one repeated task; we will show you the loop before we show you the hardware.