Four layers in a Claude decision
| Layer | Question | Evidence |
|---|---|---|
| Product | Claude app, API, Claude Code or a cloud marketplace? | Enabled surface and administrator |
| Model | Which pinned model ID and lifecycle? | Request logs and model documentation |
| Data | Training, retention, region and safety review? | Operative terms, configuration and exception |
| Workflow | Which sources, tools, actions and approvals? | Permission map and reviewer record |
What Fable 5.1 changes
Anthropic documents the API model ID claude-fable-5-1, a 1M-token context window, 128K maximum output and current list pricing. The company positions it for long-running and enterprise work. Those are product facts and vendor claims. They do not establish legal accuracy, privilege treatment or value on a firm’s matters.
What Sonnet 5.5 adds
Anthropic positions Sonnet 5.5 for well-scoped everyday work. Test its usefulness on a defined task with a source packet, acceptance criteria and a reviewer. Record the exact responding model and effort setting when comparing it with an approved baseline.
Sonnet 5.5 is compatible with zero data retention, but eligibility depends on the deployment and features. Do not extend an API retention arrangement to an ordinary Claude application or an external integration. Review the operative agreement and configuration before matter use. In the API, automatic switching is not enabled by default; customers must opt in and configure fallbacks.
See the Claude Sonnet 5.5 for Lawyers guide for candidate tasks and a reviewable contract example.
Training and retention are different fields
Anthropic’s public commercial terms say customer content is not used to train models unless the customer agrees. Its covered-model guidance separately describes a 30-day retention rule for covered models, with platform details and eligible exceptions. “No training” therefore does not mean “no retention.” Record both.
Enterprise controls still depend on the route
Anthropic describes Enterprise audit logs, SCIM and compliance tooling, while cloud marketplaces bring their own agreements and architectures. Enterprise Frontier Safeguards were announced as a phased program rather than a control every customer already has. Ask for current availability and scope instead of copying a roadmap claim into a policy.
A useful first evaluation
Select a repeatable task with public or approved data and an existing review standard. Fix the input packet and expected structure. Disable unnecessary tools. Record source fidelity, material omissions, corrections, run cost, reviewer time and export quality. Expansion requires a new decision for each data class and action boundary.
If you are past the evaluation and want to set Claude up for a practice, the Claude Projects guide for lawyers and the Claude skills vs vendor plugins decision are the two pages on this site to read next.
Where Claude sits in a legal stack
Claude may be the interface, an API model inside another product, a coding tool or one component behind a legal vendor. The system of record should remain explicit. A polished model response is not the matter file, contract repository, citation record or final approval.
| Vendor question | Operational evidence |
|---|---|
| What workflow does the product own end to end? | Inputs, transformations, tools, outputs and handoffs |
| Where is the system of record? | Repository, retention and export path |
| What happens when the model is wrong? | Detection, reviewer, rollback and correction log |
| What if the vendor or model changes? | Version pin, notice, regression test and exit path |
MCP and connectors expand the permission boundary
A connector can make a workflow more useful by reaching source systems. It also changes which data and actions are in scope. Inventory each tool, credential, destination and write permission. Start read-only, require approval for material actions and preserve the source record outside the conversation.
Ask what a legal wrapper adds
A legal product built on Claude may add licensed sources, matter controls, templates, integrations, logging, evaluation and support. Those additions can be the real product. Ask which failure modes the wrapper changes and which remain inherited from the underlying model. If the answer is only a model name, the procurement record is incomplete.
The October 28 legal artifacts gallery
On October 28, 2026, Anthropic added a gallery of ready-to-use artifacts to Claude for Legal. Anthropic names six of its own: Cite check, Jurisdiction survey matrix, Matter board, Playbook deviation board, Regulatory horizon scan and Tabular review (Anthropic post, read October 30, 2026). The same post lists four partner-built artifacts (Goodwin's AI Governance Benchmark, Harvey Intake, Foley & Lardner's Obligations Checker and PwC's Subpoena Response Tool), nine access-to-justice artifacts, and says Anthropic has added more than 60 legal connectors since May. Those counts are Anthropic's own. Harvey's post of the same day describes Harvey Intake as a launch-partner artifact that connects to Harvey through its MCP server (Harvey blog, read October 30, 2026), and Artificial Lawyer reports the same list plus a waitlisted Legal Ambassadors Program.
For a review, treat an artifact like a connector: it is a reusable tool a team can share, so record what data it reads, where its outputs are stored and who can edit it. Anthropic's own description of the workflow is that Claude completes the first pass and a lawyer applies judgment, and that each output links to its source. Read the Cite check artifact that way: a screen that points you to sources, not a verification of authority. We have not run any artifact. The sources we read give no pricing and do not say which Claude plans include the gallery, so confirm both with Anthropic before planning around it. For the Harvey side of the same launch, see our Harvey Intake and Legora Intake guide.
Claims this page does not make
- Claude preserves privilege in every configuration.
- A certification covers every feature and connector.
- A large context window makes a result complete.
- A model benchmark predicts work on a specific matter.
- A vendor confidentiality promise replaces the executed agreement.
FAQ
Does Anthropic train Claude on law-firm business data?
Anthropic’s public commercial terms say it does not train on customer content unless the customer agrees. Verify the exact product, account, feedback setting and operative agreement.
Is Claude Fable 5.1 zero data retention?
Not by default. Anthropic describes a covered-model 30-day retention rule and eligible exceptions. Confirm the approved exception and platform scope.
What are the Claude for Legal artifacts announced on October 28, 2026?
Anthropic describes them as ready-to-use tools in a legal gallery inside Claude for Legal, such as a cite check, a jurisdiction matrix and a regulatory scan, plus partner-built ones like Harvey Intake. Anthropic says Claude completes a first pass and a lawyer applies judgment. We have not tested them, and the sources we read give no pricing or plan terms.
Can Claude be used for legal research?
It can be evaluated for a defined research workflow, but every controlling authority and material factual assertion still needs independent verification.
Sources checked
- Anthropic, Introducing artifacts for legal work (Oct 28, 2026), checked 2026-10-30.
- Harvey, Harvey Intake in Claude for Legal (Oct 28, 2026), checked 2026-10-30.
- Claude Fable and Mythos 5.1 release, checked 2026-09-03.
- Fable 5.1 model overview, checked 2026-09-03.
- Covered Models retention practices, checked 2026-09-03.
- Commercial Terms, checked 2026-09-03.
- Enterprise Frontier Safeguards, checked 2026-09-03.
- Sonnet 5.5 release, checked 2026-10-03.
- API and data retention, checked 2026-10-03.
- Sonnet 5.5 model switching, checked 2026-10-03.
Free access is a different question from a deployment review. For the documented free plans and limits across vendors, see our dated table of free legal AI tools.
Operational information, not legal advice. Verify current terms, account configuration and applicable professional duties before use.