OpenAI Astra for Law: Make Every Authority Verifiable

What OpenAI launched

Astra for Law combines GPT-6 Astra with a legal search index covering US case law, statutes, regulations, court rules and administrative decisions. OpenAI says the corpus spans more than 230 million URLs, with sources added daily, and incorporates CourtListener's collection covering more than 99.9% of published US precedential case law.

Selected firms will initially access the product through Trusted Access in ChatGPT and Codex. OpenAI says it will appear as GPT-6 Astra Law and later as gpt-6-astra-law in the API. The launch also includes 26 partner-built plugins and offers Zero Data Retention for eligible firms using the API; Reuters independently confirmed the launch and its focus on research, drafting and custom legal applications.

The benchmark gain is useful, not dispositive

OpenAI tested the complete configuration on 200 US legal research questions from a private validation set of Vals AI's Legal Research Bench. At the highest reasoning effort, OpenAI reports that Astra for Law passed the overall correctness check on 54.0% of questions, versus 38.7% for GPT-6 Astra with web search alone, a 40% relative improvement. It also reports 24% more reference cases on case-law questions and up to 54% more relevant passages from the correct opinions on an audited target set.

Those are vendor-reported results on a private set, not a guarantee for a firm's matters, jurisdictions or research standards. A system can retrieve a real case and still misstate the holding, overlook controlling authority, miss later negative treatment, apply the wrong jurisdiction or fail to reflect a recent amendment. Even OpenAI's headline result leaves many questions short of the benchmark's overall correctness threshold.

Require an authority receipt for consequential work

Do not accept a citation list as proof. For each proposition that may affect advice, filing, negotiation or risk, require a structured receipt containing:

• the exact source URL, document identifier and quoted passage;
• the court or issuing body, jurisdiction, date and precedential status;
• the proposition the authority supports and any material limitation;
• subsequent history, negative treatment and the service used to check it;
• the retrieval timestamp, model and tool configuration; and
• the named lawyer who reviewed both the authority and its application to the facts.

Evaluate the workflow, not only the model

Before adoption, build a matter-representative test set that includes adverse authority, recently changed law, ambiguous jurisdiction, unpublished decisions and documents behind access controls. Score retrieval, passage fidelity, treatment checking, uncertainty, confidentiality and reviewer effort separately. Compare the full workflow with the firm's current research process, not with general web search alone.

The operating decision is also narrower than choosing one legal AI provider. Confirm which index, connector, plugin and firm source is used for each task; test permission boundaries and ethical walls; and preserve the receipt with the matter record. This complements ELYMENT AI's earlier guidance on governed legal workflows and professional accountability by making the authority itself the unit of verification.

ELYMENT AI helps organisations turn specialist AI claims into evidence, acceptance tests and human approval controls. Astra for Law may improve the research starting point. The business value arrives only when every consequential conclusion can be traced to current, controlling and independently reviewed authority.

Sources

• OpenAI, Introducing Astra for Law (17 September 2026) - Primary launch source for product scope, access, controls, legal-search coverage, integrations and vendor-reported benchmark results.
• Reuters, OpenAI launches legal-focused AI platform (17 September 2026) - Independent reporting on the launch, legal-market context, participating firms and product integrations.
• Free Law Project, CourtListener case-law coverage (18 September 2026) - First-party description of the CourtListener collection cited in OpenAI's coverage claim.
• Vals AI, legal research benchmark context (18 September 2026) - Benchmark publisher describing lawyer-authored, peer-reviewed legal research tasks and its evaluation approach.

Continue learning

• Build governed specialist legal AI workflows
• Keep professional judgement and disclosure accountable
• Test AI assurance with independent access

Frequently asked questions