Comparisons
Claude for Legal vs Astra for Law: Anthropic and OpenAI Go Direct to Lawyers
Claude for Legal vs Astra for Law: open-source plugins and MCP connectors against a 230m-URL US law index with a ZDR API, data terms, prices, the European view.
On this page
Anthropic published its legal toolkit as Apache-2.0 code that anyone can read; OpenAI, seven months later, shipped a proprietary index of 230 million URLs of US law. That is the whole difference in one sentence. As of September 2026, Claude for Legal is the usable choice for a European firm; Astra for Law is days old, US-law-centred and worth watching rather than buying until independent evaluations exist.
Two philosophies, seven months apart
Anthropic moved first. Its legal plugin for Claude Cowork was announced between 30 January and 2 February 2026, and on 3 February the shares of Thomson Reuters, RELX and Wolters Kluwer fell 13–18% in a session. On 12 May 2026 the offering became Claude for Legal: twelve practice-area plugins (Commercial, Corporate, Employment, Privacy, Product, Regulatory, AI Governance, IP, Litigation, Legal Clinic, Law Student and a Legal Builder Hub), with 90-plus agents and skills and 16 connectors at launch. The public repository has since grown to 150-plus skills, 13 scheduled agents and 20-plus MCP connectors, with 9,500 GitHub stars, under an Apache 2.0 licence. It runs in Claude Cowork (merged into the main Claude app on 17 September 2026), Claude Code, Managed Agents and Microsoft 365, and paid Claude customers get it at no extra charge.
OpenAI's answer took longer and looks different. Jason Boehmig, co-founder of the contract platform Ironclad, joined OpenAI in June 2026 to lead legal. GPT-6 Astra went public to paid users on 4 September 2026. Then on 17 September OpenAI announced Astra for Law: GPT-6 Astra configured for legal work, a Legal Search Index of 230 million-plus URLs of US law built with CourtListener and the Free Law Project, 26 partner plugins (Thomson Reuters, Harvey, Legora, iManage, Intapp and DeepJudge are the named ones) plus 47 community plugins, a zero-data-retention API and a "Trusted Access" programme, with Latham & Watkins, Sullivan & Cromwell, Ropes & Gray and Cooley as design partners.
Artificial Lawyer's Richard Tromans framed the contest as "the battle for centrality": where does a lawyer's work primarily live, and which interface owns the day. The detailed walkthroughs of the Claude for Legal plugins and of Astra for Law cover each in depth; this page sets them against each other.
What ships in each box
| Component | Claude for Legal (Anthropic) | Astra for Law (OpenAI) |
|---|---|---|
| Launched | 3 Feb 2026 (first plugin); 12 May 2026 (12 plugins) | 17 Sep 2026 |
| Model | Fable 5.1, Opus 5, Sonnet 5 as the account allows | GPT-6 Astra (public 4 Sep 2026; off by default in ChatGPT Enterprise until an admin enables it) |
| Legal content | None of its own; citations not checked against a connected tool are tagged [verify] | Legal Search Index, 230m-plus URLs of US law (CourtListener / Free Law Project) |
| Practice tooling | 12 open-source plugins, 150-plus skills, 13 scheduled agents; cold-start interview writes a practice profile | 47 community plugins written by lawyer "power users"; design-partner builds (S&C negotiating-playbook analyser, Cooley IPO-prep tool) |
| Integrations | 20-plus MCP connectors: iManage, NetDocuments, DocuSign, Ironclad, Relativity, Everlaw, Box, Datasite, Thomson Reuters CoCounsel, LexisNexis, Harvey, Midpage, Trellis, Definely | 26 partner plugins; four named publicly; LawSites asked OpenAI for the full list and "got no answer" |
| Guardrails | Every output "a draft for attorney review"; send gates on demands and takedowns; review gate before board consents | "Trusted Access"; OpenAI safety testing reported GPT-6 Astra exceeding authorised targets 0% of the time versus 48% for GPT-5.6 Sol |
| Licence | Apache 2.0; you can read, fork and edit the plugins | Proprietary |
| Surfaces | Claude app (Cowork), Claude Code, Managed Agents, Microsoft 365 | ChatGPT (Business, Enterprise), API |
| Price | Included with paid Claude plans (Pro $20, Max $100/$200, Team $25/$125 per seat, Enterprise custom); API $10/$50 per million tokens for Fable 5.1, $5/$25 for Opus 5 | Not separately published; GPT-6 Astra API $10/$50 per million tokens; ChatGPT Business $25 per seat, Enterprise custom |
Two things stand out. The first is that Anthropic sells no legal content at all: the plugins know how to run a diligence grid "every cell cited" or a DSAR within its statutory clock, but the authorities come from whatever research tool you connect. GC AI's early review named three gaps: persistent context, citation verification and real case-law access. OpenAI's index is a direct answer to the third gap, for US law. The second is that OpenAI's box is closed. You cannot read how Astra for Law's plugins gate a filing; you can read every line of Anthropic's FRE 408 gate on demand letters.
Data terms: the section that decides it for a regulated firm
Neither vendor trains on business-tier data. The retention details differ, and they matter more than the marketing.
Anthropic's retention documentation is explicit: Team, Enterprise and API content is never used for training, and API content is not retained by default. The exceptions are the strongest models. Fable 5 and 5.1 (and their classifier-free siblings, Mythos) are "Covered Models" that require 30-day retention and are not eligible for zero data retention without express authorisation. The Claude Team and Enterprise chat interfaces are not ZDR-eligible either; Claude Code under Enterprise is. Session transcripts captured by the Compliance API are kept six years by default. Consumer Free, Pro and Max users had to choose whether their chats train models, with opted-in data kept five years.
OpenAI's enterprise privacy terms say Business, Enterprise, Edu and API content is not used for training by default; API content is retained 30 days for abuse monitoring; zero data retention is contractual. Astra for Law advertises a ZDR API, which for a firm that has been told by its bank clients to insist on ZDR is a genuine advantage over Fable. Consumer ChatGPT trains unless the user opts out, and the In re OpenAI copyright litigation showed that 20 million de-identified consumer logs can be ordered produced.
The practical reading: a firm can get zero retention from OpenAI's API tier, and from Anthropic's API tier on Opus 5 or Sonnet 5 but not on Fable 5.1. Whichever you pick, the confidentiality checklist and the zero data retention entry explain how to ask for it in writing.
Benchmarks: the 54% that started an argument
OpenAI's launch number was that Astra for Law scored 54% on the Vals Legal Research Bench, against 38.7% for GPT-6 with plain web search. The figure is OpenAI-reported, and it drew a 655-comment Hacker News thread in which one commenter asked how a product that "passed the evaluation's overall correctness check on 54.0% of questions" could be sold at all, and another coined "Lawslop". Both reactions miss the context: in Vals' October 2025 study, ChatGPT Team scored 80% on a weighted accuracy metric against a 69% lawyer baseline, and different Vals benchmarks score different things. The benchmarks guide untangles them.
The independent picture is mixed. On the Vals Index of 15 September 2026 (58 models), Claude Fable 5.1 is first, Claude Opus 5 second and GPT-6 Astra third. On Harvey's BigLaw Bench, OpenAI leads: GPT-5.6 Sol 92.7% and GPT-5.5 91.7% against Claude Opus 4.7's 90.9%. On Harvey's stricter all-criteria-pass LAB, an aggregator leaderboard on 18 September 2026 has Fable 5 at 13.3%, Opus 5 at 11.7% and GPT-5.5 at 2.1%. On HAQQ's June 2026 test, GPT-5.5 had the lowest citation-hallucination rate (3%) while Claude Opus 4.8 won most tasks overall. GC AI's vendor-run in-house bench put GPT-5.5 at 79.8% and Opus 4.7 at 68.4%. The Claude vs ChatGPT comparison works through these task by task; the short version is that the winner depends on who wrote the test, and that nobody has yet tested Astra for Law independently.
Who has signed up, and what they built
Anthropic's legal references are a year deep. Freshfields signed a multi-year deal on 23 April 2026 to put Claude in front of 5,700 staff across 33 offices and reported roughly 500% usage growth in six weeks. More than 20,000 people registered for Anthropic's April 2026 legal webinar, and Anthropic says legal professionals are the most engaged Cowork users of any function. Thomson Reuters rebuilt CoCounsel Legal on the Claude Agent SDK, Harvey turned on Opus 5 on 24 July 2026 and Fable 5.1 on 1 September, and Microsoft's Word Legal Agent requires Anthropic to be enabled as a sub-processor.
OpenAI's references are a week deep but well chosen. Latham & Watkins is a design partner, Sullivan & Cromwell built a negotiating-playbook analyser and Cooley an IPO-preparation tool, and the 47 community plugins came from lawyer power users. What is missing is the long tail: the full list of 26 partner plugins, pricing, non-US availability and any deployment story with usage numbers.
What it means for Harvey and Legora
Both labs now sit above and beside the legal-AI platforms at once. Harvey is a connector inside Claude for Legal and a partner plugin inside Astra for Law; Legora is a partner plugin in Astra and shipped its own ChatGPT Enterprise plugin on 17 September 2026. Harvey runs a multi-model stack and released its own post-trained open-weight model, Tenet, in research preview on 20 August 2026, which is a hedge against exactly this moment. Harvey's CEO has conceded that "long term we would end up competing with the model companies", and Law.com asked in July 2026 who is not using Harvey or Legora, and why.
The vendors' defence is grounding, governance and all-inclusive pricing: a Reddit thread on rejecting Harvey and Legora in August 2026 recorded the strongest pro-vendor argument as "it's anyone's guess who will have the best AI models in 5 years", and the strongest counter as a 100-lawyer firm's technology committee choosing Claude Enterprise over both. The Harvey vs Legora page covers that fight; the point here is that whichever platform you buy, the model underneath now comes from one of these two labs, and the labs are selling to you directly.
For a mid-sized European firm
The US-law index is the headline feature of Astra for Law and it is close to irrelevant for a dispute under the ABGB or a German employment matter. Claude for Legal's plugins are also US-centric in their defaults (FMLA and CFRA leave trackers, DMCA notices, FRE 408 gates), but they are open source, and the cold-start interview exists precisely so a firm can rewrite the practice profile for its own jurisdiction and playbooks. A knowledge lawyer with no coding background can read a skill file and change it.
The regulatory overlay favours the same answer for now. Both vendors signed the EU GPAI Code of Practice, and the EU AI Act puts a deployer's AI-literacy and transparency duties on the firm whichever it uses. Under GDPR Art. 28 the firm needs a processing agreement and a retention position it can defend, and the GDPR guide explains what to ask for. The Austrian bar's four rules (client data on EU servers, disclosure to clients, the lawyer keeps final responsibility, no client data for training) and BRAK's guidance on § 43a BRAO are satisfiable on Anthropic's business tiers today, with Fable's 30-day retention as the one item to document; Astra for Law's European terms are not yet published.
What is still unknown about Astra for Law, as of 18 September 2026:
- Pricing, seat minimums and whether the Legal Search Index is included or metered.
- Availability and data residency outside the United States.
- The full list of 26 partner plugins.
- Any independent benchmark or hallucination measurement.
- How "Trusted Access" is granted and what it gates.
- Whether the "recurrent depth" technique that obscures some of GPT-6 Astra's reasoning affects explainability of legal outputs.
A week-one evaluation for either
Do not pilot on a live matter. Run this on a closed file with the names changed.
- Enable the business tier (Claude Team or Enterprise; ChatGPT Business or Enterprise, with GPT-6 Astra switched on by an admin) and get the retention terms in writing.
- Connect one document source through MCP or a partner plugin, following the DMS connection guide, and confirm that ethical walls hold by testing with a walled user.
- Run the cold-start interview (Claude) or the equivalent onboarding, and give the tool your NDA playbook. Module 4 of the AI Lab for Lawyers course sets up these plugins live, including the interview and a first scheduled agent.
- Run the same prompt in both:
You are assisting a supervising lawyer. Review the attached NDA against the
attached playbook. For each departure from the playbook: quote the clause,
state the playbook position, propose replacement wording as a tracked change,
and grade the risk RED, YELLOW or GREEN. Cite no authority you have not
retrieved from a connected source; mark anything unverified as [verify].
Do not send, file or share anything. Stop and ask before any action that
would leave this workspace.
- Score each tool on the departures it found, the departures it invented, and whether it honoured the stop instruction. Open every citation.
- Decide on the evidence, not on the launch coverage, and re-run the test in a quarter when Astra for Law has terms, prices and an independent evaluation.
Whichever wins your test, the Model Context Protocol is the piece that keeps you portable: a DMS connector that works with Claude, ChatGPT, Copilot, Harvey and Legora means the next lab launch is a switch, not a migration.
Frequently asked questions
What is the difference between Claude for Legal and Astra for Law?
Claude for Legal is Anthropic's open-source set of practice-area plugins, skills and MCP connectors that run inside Claude Cowork, Claude Code, Managed Agents and Microsoft 365, at no extra charge to paid Claude customers. Astra for Law is OpenAI's GPT-6 Astra configured for legal work, with a proprietary Legal Search Index of 230 million-plus URLs of US law, 26 partner plugins and a zero-data-retention API. One is code you can inspect; the other is a hosted service.
Is Astra for Law better than Claude for Legal?
Nobody can say yet. Astra for Law launched on 17 September 2026. OpenAI reports 54% on the Vals Legal Research Bench against 38.7% for GPT-6 with plain web search, but that is a vendor figure. On the independent Vals Index of 15 September 2026, Claude Fable 5.1 and Opus 5 sit first and second and GPT-6 Astra third. On Harvey's BigLaw Bench, OpenAI models lead. Run both on your own matters before believing either side.
Does Claude for Legal or Astra for Law use my data for training?
Neither trains on business-tier data. Claude Team, Enterprise and API content is never used for training; ChatGPT Business, Enterprise and API content is not used by default. The differences are in retention: OpenAI keeps API content 30 days for abuse monitoring unless zero data retention is agreed, and Astra for Law advertises a ZDR API. Anthropic's Fable 5 and 5.1 require 30-day retention and are excluded from ZDR without express authorisation.
Should a European law firm use Claude for Legal or Astra for Law?
As of September 2026, Claude for Legal, if either. Its plugins are US-centric but open source, so a firm can adapt them to Austrian, German or English practice through the cold-start interview and its own playbooks, and MCP connectors reach iManage and NetDocuments under existing permissions. Astra for Law's Legal Search Index covers US law, its European availability and terms are not yet published, and no independent evaluation exists. Revisit in a quarter.
What does Astra for Law mean for Harvey and Legora?
Both are among Astra for Law's 26 partner plugins and both are connectors inside Claude for Legal, so the labs supply and compete with them at once. Legora shipped a ChatGPT Enterprise plugin on 17 September 2026; Harvey runs Claude Opus 5 and Fable 5.1 alongside OpenAI models and has released its own open-weight model, Tenet. The vendors' defence is grounding, governance and all-inclusive pricing; the labs' attack is price and model quality.
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
Claude for Legal Plugins Explained: The 12 Practice-Area Plugins and 20+ Connectors
Claude for Legal gives paid Claude users 12 open-source practice-area plugins, 150+ skills and 20+ connectors to iManage, NetDocuments and CoCounsel.
8 min read
Tools
Astra for Law: What OpenAI's Legal Offering Actually Includes
Astra for Law, announced 17 September 2026, pairs GPT-6 Astra with a 230-million-URL Legal Search Index, 26 partner plugins and a zero-retention API.
8 min read
Comparisons
Claude vs ChatGPT for Lawyers (2026): Which Should You Use?
Claude vs ChatGPT for lawyers in 2026, compared task by task: drafting, research, agents, data terms by plan, prices and benchmarks, plus a pick by firm type.
10 min read
Comparisons
Harvey vs Legora: The Two Legal AI Platforms Compared
Harvey vs Legora in September 2026: $15.5bn vs $5.6bn valuations, ARR, clients, Workflow Agents vs Agentic OS, models, pricing minimums, geography and who should pick which.
10 min read
Glossary
Model Context Protocol (MCP) — Definition for Lawyers
Model Context Protocol (MCP) is the open standard that lets Claude, ChatGPT and Copilot read your DMS or docket feed through one connector. What a server exposes and what to check.
4 min read
Glossary
Zero Data Retention (ZDR) — Definition for Lawyers
Zero data retention means the AI vendor stores nothing after replying. Which plans offer ZDR, the Covered Models exception, and why it matters for privilege and GDPR.
5 min read