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.
On this page
Astra for Law is OpenAI's legal product, announced on 17 September 2026 and reported on 18 September. It configures GPT-6 Astra for legal work and adds a Legal Search Index of 230 million-plus URLs of US law built with CourtListener, 26 partner plugins, 47 community plugins, a zero-data-retention API and a Trusted Access programme. Design partners are Latham & Watkins, Sullivan & Cromwell, Ropes & Gray and Cooley. It is one day old, and almost everything about its accuracy remains unverified.
What was announced on 17 September
OpenAI's announcement had five parts (LawSites; Artificial Lawyer; OpenAI):
- A configured model. GPT-6 Astra tuned for legal drafting, analysis and research, not a separate legal model.
- A Legal Search Index. More than 230 million URLs of US law, assembled with CourtListener, the Free Law Project's non-profit case-law database. CourtListener is reported to cover 99.9% of published US precedent; treat that figure as reported rather than audited.
- Plugins. 26 partner plugins from vendors including Thomson Reuters, Harvey, Legora, iManage, Intapp and DeepJudge, plus 47 community plugins reportedly written by lawyer "power users". Legora shipped its ChatGPT Enterprise plugin the same day.
- Data terms. A zero-data-retention API and a "Trusted Access" programme for firms.
- Design partners. Latham & Watkins, Sullivan & Cromwell, Ropes & Gray and Cooley. Law.com reports Sullivan & Cromwell built a negotiating-playbook analyser on it and Cooley an IPO-preparation tool (Legaltech News).
Behind it is a hire: Jason Boehmig, co-founder of the CLM vendor Ironclad, joined OpenAI in June 2026 to lead legal. That is the clearest signal of intent. Ironclad has a reported $200 million in ARR and Boehmig knows exactly which contract workflows in-house teams pay for.
The model underneath: GPT-6 Astra
Astra for Law inherits everything about GPT-6 Astra, which reached paid ChatGPT users on 4 September 2026 after a limited preview on 3 September (9to5Mac). The facts that matter to a lawyer:
| Attribute | GPT-6 Astra (as of 18 Sep 2026) |
|---|---|
| API price | $10 per million input tokens, $50 output, $1 cached |
| Context window | About 1.05 million tokens; 128K output |
| Knowledge cutoff | 30 April 2026 |
| Enterprise default | Off until an administrator enables it |
| Reasoning visibility | "Recurrent depth" technique obscures some or all of the model's reasoning |
| Vals Index rank (15 Sep 2026, 58 models) | Third, behind Claude Fable 5.1 and Claude Opus 5 |
| ARC-AGI-3 | 99.9% with tools, against 30% for Opus 5 |
Two of those rows deserve a pause. The Enterprise default means the general counsel or IT administrator, not the associate, decides whether the firm gets the model at all. And the obscured reasoning means that when Astra for Law produces a research memo, you may not be able to see how it got there; the how LLMs work guide explains why that matters more for legal analysis than for code.
The reported knowledge cutoff of 30 April 2026 is also why the Legal Search Index exists. Without live retrieval, a model that stopped reading in April cannot know about a judgment from June. With it, the question becomes whether the retrieval is right, which brings us to the benchmark.
The one number, and why it is not enough
OpenAI reported at launch that Astra for Law answered 54% of a 200-question set from the Vals Legal Research Bench correctly, against 38.7% for GPT-6 Astra using ordinary web search. That is a vendor-run figure on a subset, and it is the only accuracy number that exists. For context, on Vals' full Legal Research Bench of 15 September 2026 the top three were Meta's Muse Spark 1.3 Max, Claude Opus 5 and Claude Fable 5.1 (Vals), and in Vals' October 2025 study ChatGPT with web search already scored about 80% on a different, older question set, above a 71% lawyer baseline. Different sets, different scoring; the benchmarks guide explains why these numbers cannot be lined up.
What can be said is that retrieval reduces fabricated cases and does not eliminate misgrounded ones. In the HAQQ benchmark of June 2026, 24% of 3,000 answers across leading models cited or misapplied law that did not support the claim, and a University of Pennsylvania study in April 2026 found search-augmented models and deep-research agents still fabricated 3–13% of URLs (arXiv). A 230-million-URL index is a better haystack; it is not a reader. The AI legal research page sets out the verification routine that applies regardless of vendor.
There is also an uncomfortable footnote in the design-partner list. Sullivan & Cromwell apologised to Bankruptcy Judge Martin Glenn in April 2026 for AI-generated fake citations in an 8 April filing, which the firm's review "did not identify" (Bloomberg Law). Five months later the firm is building on Astra for Law. Whether that reflects confidence in the new index or a determination not to repeat the episode, the lesson is the same: the firm's process, not the vendor's, caught nothing the first time.
Data terms: ZDR, Trusted Access and the consumer trap
OpenAI's stated data terms for Astra for Law are a zero-data-retention API and the Trusted Access programme. The general ChatGPT rules still frame them (OpenAI):
| Plan | Trains on your content? | Retention |
|---|---|---|
| Free, Plus, Pro | Yes, unless the user opts out | Standard |
| Business ($25 per seat monthly, $20 annual), Enterprise, Edu | No by default | Standard, with admin controls |
| API | No by default | 30 days for abuse monitoring |
| API with ZDR | No | Contractual zero retention |
Two court decisions make the consumer row dangerous for lawyers. In United States v. Heppner (S.D.N.Y., opinion 17 February 2026), Judge Rakoff held that a defendant's research memos generated with consumer Claude were neither privileged nor work product, because the terms allowed logging and disclosure and "Claude is not an attorney"; the same logic applies to a Plus account. And in In re OpenAI Copyright Litigation (S.D.N.Y.), 20 million de-identified ChatGPT logs were ordered produced in November 2025, affirmed on 5 January 2026: logs are discoverable ESI. The privilege and AI page covers both; the practical rule is that Astra for Law belongs on Enterprise, Business or a ZDR API contract, and that "Trusted Access" needs to be read as a document, not a label. The zero data retention entry explains what to ask for.
How it compares with Claude for Legal
The full comparison is its own page; the structural difference fits in a table.
| Claude for Legal (12 May 2026) | Astra for Law (17 Sep 2026) | |
|---|---|---|
| What it is | 12 open-source practice-area plugins, 150+ skills, 20+ connectors | A configured model plus a proprietary US legal index plus partner plugins |
| Licence | Apache 2.0; readable and editable | Proprietary |
| Research grounding | Via connectors (CoCounsel, LexisNexis, CourtListener, Trellis); unverified citations tagged [verify] | Built-in Legal Search Index of 230 million-plus URLs |
| Runs in | Cowork, Claude Code, Managed Agents, Microsoft 365 | ChatGPT (Enterprise off by default), API |
| Price | Included with paid Claude plans | Not separately published as of 18 Sep 2026 |
| Independent evaluation | Harvey LAB: Fable 5 13.3% strict all-pass (18 Sep 2026) | None yet; vendor figure of 54% on a 200-question subset |
| Named firms | Freshfields (5,700 staff, 33 offices) | Latham, Sullivan & Cromwell, Ropes & Gray, Cooley |
One is open and connector-driven, the other integrated and days old. Harvey and Legora, meanwhile, will offer Astra through their own platforms via the API, so the choice for a firm already on one of them is less stark than the headlines suggest; the Legora review covers its ChatGPT Enterprise plugin.
What nobody knows yet
- Whether the Legal Search Index covers anything outside US law, and when. For a European firm this is the first question.
- Pricing, seat minimums and whether Astra for Law is a ChatGPT Enterprise add-on or a separate contract.
- EU data residency and sub-processor lists for the index and the plugins.
- Any independent hallucination or misgrounding rate.
- How the 47 community plugins are vetted. Anthropic's equivalent, the legal-builder-hub, ships with an injection scan and licence gate; nothing comparable has been described for Astra.
- Product longevity. OpenAI retired the ChatGPT Atlas browser on 9 August 2026 after 292 days, removed the standalone ChatGPT agent in early August (reported by a single source) and announced on 3 June 2026 that Agent Builder and Evals close on 30 November 2026 (OpenAI community notice). None of that means Astra for Law will follow, but it argues for contracting for data and plugin export from the start.
A cautious pilot plan
If your firm is already on ChatGPT Enterprise, the pilot is cheap: ask the administrator to enable GPT-6 Astra for a named group, request Trusted Access terms in writing, and run the same known-answer research questions you would put to any tool, including a misgrounding trap and a multi-jurisdiction question, against your current research platform. Record the results in a matter-independent log. If you are not on Enterprise, do not pilot on Plus accounts; the ChatGPT for lawyers page explains what to switch off before anyone types a client name. Give it sixty days for independent evaluations to appear before committing budget, and run the questions in the vendor due diligence guide in the meantime. Module 6 of the AI Lab for Lawyers course covers how to assess a brand-new legal AI product against the risk, regulatory and confidentiality tests before it touches a matter.
Frequently asked questions
What is Astra for Law?
Astra for Law is OpenAI's legal offering announced on 17 September 2026. It configures GPT-6 Astra for legal work and adds a Legal Search Index of more than 230 million URLs of US law built with CourtListener and the Free Law Project, 26 partner plugins from vendors such as Thomson Reuters, Harvey, Legora, iManage, Intapp and DeepJudge, 47 community plugins, a zero-data-retention API and a Trusted Access programme.
How accurate is Astra for Law for legal research?
As of 18 September 2026 the only published figure is OpenAI's own: 54% correct on a 200-question set from the Vals Legal Research Bench, against 38.7% for GPT-6 Astra using plain web search. No independent evaluation exists yet. On the separate Vals Index across all tasks GPT-6 Astra ranked third behind Claude Fable 5.1 and Opus 5 on 15 September 2026.
Is Astra for Law available in Europe?
OpenAI has not published regional availability, EU data residency or non-US content coverage for Astra for Law as of 18 September 2026. The Legal Search Index is described as US law. GPT-6 Astra itself is available to paid ChatGPT users and through the API, but is off by default for Enterprise tenants until an administrator enables it.
How does Astra for Law compare with Claude for Legal?
Claude for Legal, expanded on 12 May 2026, is a set of 12 open-source practice-area plugins under Apache 2.0 that runs on any paid Claude plan and connects to research tools through partners. Astra for Law bundles a model, a proprietary US legal index and partner plugins inside OpenAI's platform. One is open and connector-driven, the other is integrated and days old; a fair comparison needs independent testing that does not yet exist.
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
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.
11 min read
Tools
ChatGPT for Lawyers: GPT-6 Astra, ChatGPT Work and What to Switch Off
ChatGPT for lawyers as of September 2026. Which plans train on your chats, the 20-million-log discovery order, GPT-6 Astra, ChatGPT Work and a firm setup.
8 min read
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
Guides
Legal AI Benchmarks Explained: Vals, BigLaw Bench, LAB, HAQQ and LegalBench
Legal AI benchmarks explained: what Vals, BigLaw Bench, Harvey LAB, HAQQ and LegalBench measure, the September 2026 leaders, and how to run your own evaluation.
10 min read
Regulation & Ethics
Attorney–Client Privilege and AI: Heppner, Warner, the NYT Logs and Public Chatbots
Consumer chatbot memos are not privileged. US v. Heppner, Warner v. Gilbarco and the 20-million-log OpenAI order, plus rules that keep privilege intact.
10 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