GPT-6 Astra Law: OpenAI's Astra for Law Turns Legal Research Into a Measurable Contest
OpenAI has released Astra for Law, a version of GPT-6 Astra fitted with legal-specific settings, tools, and a dedicated legal search index, and the company is backing it with a head-to-head benchmark against its own general-purpose model.
OpenAI has released Astra for Law, a version of GPT-6 Astra fitted with legal-specific settings, tools, and a dedicated legal search index, and the company is backing it with a head-to-head benchmark against its own general-purpose model.
A model built around a legal search index rather than general web search
Astra for Law is aimed at law firms and legal technology companies, and its core addition is a legal search index covering more than 230 million URLs of U.S. case law, statutes, regulations, court rules, and administrative decisions. OpenAI adds sources to this index daily, and it includes a CourtListener collection that covers more than 99.9% of published U.S. precedential case law. That scope matters because it defines the ceiling of what the model can retrieve when a lawyer asks it to locate authority on a given question. The system also carries custom instructions for legal analysis and writing, layered on top of the underlying GPT-6 Astra model. Whether that combination changes outcomes in practice is the question OpenAI tries to answer with numbers, not adjectives.
The numbers OpenAI is using to make its case
OpenAI ran Astra for Law against the Vals AI Legal Research Bench and reports an overall correctness rate of 54.0% of questions, compared with 38.7% for GPT-6 Astra using web search alone. That is a 40% relative improvement on the same evaluation. On case-law-focused questions, Astra for Law found 24% more reference cases than the web-search-only version of the base model, and on an audited target-passage set it retrieved up to 54% more relevant passages from correct court opinions. These are direct, matched comparisons between the legal-tuned system and its own general-purpose sibling, run on the same benchmark, which gives the figures more weight than a standalone accuracy score would carry on its own. They also set a baseline that outside evaluators can test independently, since Vals AI is a named, external benchmark rather than an internal metric.
Where it lives, who can reach it, and under what handling terms
In ChatGPT, the model appears in the model picker as GPT-6 Astra Law, and in the API it is addressable as gpt-6-astra-law, with OpenAI stating that API availability is planned rather than immediate. At launch, access runs through Trusted Access in ChatGPT and Codex, and it is being offered initially to selected law firms rather than opened broadly. For firms that do get in, OpenAI specifies concrete data terms: Zero Data Retention applies to API usage, and ChatGPT Enterprise usage for eligible firms is excluded from human review by default. Those are the only handling commitments OpenAI has stated, and they apply specifically to eligible firms under this program rather than to Astra for Law as a general offering.
An ecosystem of plugins arriving alongside the core model
Astra for Law launched with 26 partner-built plugins and 9 community plugins, and the community plugins collectively include 47 custom skills. OpenAI has not detailed what each plugin does or which vendors built them, so the practical shape of that ecosystem is still an open question rather than a settled feature set. What did ship as generally available at launch is ChatGPT for Word, released alongside the legal model. Together, the index, the benchmark results, the Trusted Access rollout, and the plugin count describe a product that OpenAI is positioning less as a chatbot with a legal skin and more as an evaluated research tool — one whose real test will be how selected firms use it once Trusted Access opens their access and the promised API availability actually arrives.
Specs
- Base model: GPT-6 Astra
- API model ID: gpt-6-astra-law
- Legal search corpus: More than 230 million URLs
- Partner-built plugins: 26
- Community plugins: 9
- Custom skills: 47
Sources
- OpenAI — OpenAI