The news
Primary: OpenAI, 17 September 2026, Introducing Astra for Law.
Astra for Law is GPT-6 Astra combined with a legal search index and instructions for legal analysis and writing. OpenAI frames it as a foundation for law firms and legal-technology companies to build products and workflows around their own expertise.
Legal search index (OpenAI's description): U.S. case law, statutes, regulations, court rules, and administrative decisions across a corpus of more than 230 million URLs, with sources added daily. OpenAI's work with Free Law Project (CourtListener) brings a case-law collection covering more than 99.9% of published U.S. precedential case law into the research experience. The index complements licensed content and specialist products from providers such as Thomson Reuters.
Bench (OpenAI-reported, private validation set): Vals AI Legal Research Bench, 200 U.S. legal research questions. At the highest reasoning effort for both systems:
| Metric | Astra for Law | GPT-6 Astra + web search |
|---|---|---|
| Overall correctness | 54.0% | 38.7% (about 40% relative improvement) |
| Case-law-focused: reference cases | 24% more than web-search baseline | baseline |
| Relevant passages from correct opinions | up to 54% more at same reasoning effort | baseline |
Availability: initially offered to selected law firms through Trusted Access in ChatGPT and Codex; coming soon to the API. Model picker label: GPT-6 Astra Law. API id: gpt-6-astra-law.
API customers including Harvey and Legora can build on Astra for Law. OpenAI also launches 26 partner-built ecosystem plugins connecting ChatGPT to specialist legal tools; Relativity and Clio are named among the tools firms already use. Community plugins and ChatGPT for Word general availability are also on the page.
Attributed preview quotes on the primary page include John Savva (Partner, Sullivan & Cromwell) on research depth and sensitivity to authority, and Niko Grupen (Head of Applied Research, Harvey) on grounding, citation precision, and advisory guidance. Firm build examples named on the page: Sullivan & Cromwell (agreement analyzer), Ropes & Gray (deal diligence), Cooley (GO Public capital-markets tooling). Governance collaboration: Latham & Watkins on information permissions, ethical walls, client instructions, and firm oversight (Michael Rubin quoted as Chair of Latham's AI Strategy Committee).
Who is bound
No statute or regulator mandate attaches to this launch. The announcement binds OpenAI to the product terms it publishes for Trusted Access, ChatGPT Enterprise, and the API.
Operators who take the product:
- Selected law firms on Trusted Access (ChatGPT and Codex) for professional legal work under lawyer supervision.
- API builders (Harvey, Legora, and others) once
gpt-6-astra-lawships to the API. - Plugin and partner vendors connecting through the 26 partner plugins and community plugins.
Privacy controls OpenAI states for eligible firms: Zero Data Retention (ZDR) on the API; ChatGPT Enterprise usage excluded from human review by default. Broader firm governance design is described as work in progress with Latham & Watkins.
Professional-conduct duties (competence, confidentiality, supervision, client instruction) remain with the lawyers and firms under applicable state bars. OpenAI does not claim the product replaces licensed counsel.
What's new
- A frontier-model + legal-index configuration sold as a named SKU (
gpt-6-astra-law/ GPT-6 Astra Law), rather than generic web search alone. - A published head-to-head on Vals AI's Legal Research Bench private set with concrete correctness and retrieval deltas versus GPT-6 Astra + web search.
- CourtListener / Free Law Project case-law coverage wired into the research path at the scale OpenAI cites (more than 99.9% published U.S. precedential case law; more than 230M URLs).
- Trusted Access packaging with ZDR and default exclusion from human review for eligible ChatGPT Enterprise usage, plus named governance work with Latham & Watkins.
- Twenty-six partner plugins at launch, including named practice tools (Relativity, Clio) and a path for API partners such as Harvey and Legora.
What it does not settle
- Whether 54.0% overall correctness on a 200-question private validation set generalizes to live firm matters, other jurisdictions, or adversarial document sets. Outside that bench: UNKNOWN. Independent public replication of the Vals private-set scores: UNKNOWN.
- Billing, seat limits, and which firms qualify for Trusted Access. Not specified in the primary beyond "selected" and "eligible."
- Exact API GA date. OpenAI says "coming soon."
- How licensed Thomson Reuters (and other) content interacts with the open/CourtListener index inside a given matter workflow. Complements, per OpenAI; conflict and licensing detail: UNKNOWN from this page alone.
- Malpractice, privilege, or bar-ethics outcomes when lawyers rely on Astra for Law outputs. The product page does not settle those questions.
- Whether the demo prompts on the page (including comparisons to other named models) are representative of production quality. Treat as marketing evaluation context.
What to do now
- If you run legal AI procurement: request Trusted Access criteria, ZDR scope, human-review defaults, and data-retention terms in writing before pilot.
- If you build on the API (Harvey/Legora-class or internal): plan for model id
gpt-6-astra-law; confirm rate limits and ZDR enrollment when GA lands. - If you evaluate research quality: re-run your own golden set against Vals-style criteria; do not treat the 54.0% figure as a substitute for matter-specific validation.
- If you already use Relativity, Clio, iManage, or HighQ: map which of the 26 plugins touch matter files, time entry, or DMS write-back, and who approves those connectors.
- Keep a human cite-check step for authorities and holdings. OpenAI's own framing puts the lawyer in the loop to examine authorities.
- Watch for API GA and any change to the CourtListener / licensed-content mix. Those dates and license terms are UNKNOWN until OpenAI or partners publish them.