OpenAI expands Astra for Law to all Enterprise customers
OpenAI is opening its specialized legal AI configuration to all Enterprise users while upgrading the underlying model and expanding its global search index.
The short version
- Astra for Law is a configuration of the GPT-6.2 Astra model featuring specific legal instructions and tools.
- The legal search index grew by 34% to cover over 310 million URLs, including data from the U.K., Australia, and Canada.
- The model achieved a 20.8% success rate on the LAB-AA benchmark, outperforming the base model and other competitors.
- Astra for Law maintains one of the lowest hallucination rates in its class, averaging 0.03 material hallucinations per assignment.
Why it matters
This move signals a shift from exclusive, pilot-based legal AI to a broader enterprise standard. For legal professionals, it means more reliable, specialized tools for drafting and research are becoming accessible to larger firms.
Broadening access for firms
OpenAI is expanding the reach of Astra for Law, a version of its GPT-6 Astra model tailored specifically for the legal industry. While the tool was initially limited to a small group of "Trusted Access" firms, it is now available to every ChatGPT Enterprise customer. Users can request access through the admin console or apply for API access to integrate the tool into their own software.
This transition follows a period of high-profile competition in the legal AI space. Other major tech companies have recently launched their own legal-specific offerings, including Anthropic's Claude for Legal, Google's Gemini Enterprise for Legal, and Microsoft's Copilot for Legal. OpenAI's move aims to solidify its position as a primary provider for professional legal work.
The company describes the tool not as a brand-new model, but as a specialized configuration. It uses the base GPT-6.2 Astra model combined with specific legal instructions, specialized tooling, and a dedicated search index. This approach allows the AI to better handle the nuances of legal research, weighing different authorities, and drafting comprehensive answers.
Upgraded model and search index
The latest update moves the legal configuration onto the newer GPT-6.2 Astra base model. This upgrade provides a significant boost in performance over the previous version. According to the company, the new configuration also includes updated instructions for researching questions and writing legal answers.
A major component of this release is the expansion of the legal search index. The index originally covered over 230 million URLs of U.S. case law, statutes, and regulations. It has since grown by 34%, adding more than 80 million new URLs that include material from Canada, Australia, and the United Kingdom.
The expanded index now covers more than 99.9% of published U.S. precedential case law. By providing a broader scope of data, the tool aims to give legal professionals a more comprehensive view of the law across different jurisdictions. This data draws heavily on the CourtListener database from the nonprofit Free Law Project.
Benchmark performance and accuracy
OpenAI claims that Astra for Law achieved the top score on a benchmark of legal agent work developed by Harvey. In the LAB-AA test, which covers 120 assignments across 24 areas of law, Astra for Law earned full credit on 20.8% of the tasks. This outperformed the base GPT-6.2 Astra model, which scored 15.0%, and the Muse Wonder Full model, which scored 13.1%.
The scoring system for these assignments is strict, requiring the AI to meet every requirement without a material hallucination. Analysts have noted that the gap between the legal configuration and the base model is significant. It represents the difference between finishing one in five assignments cleanly versus one in seven.
Accuracy remains a primary focus for the legal configuration. The model averaged 0.03 material hallucinations per assignment, which tied for the lowest rate in the comparison. In contrast, Anthropic’s Claude Opus 5.5 averaged 0.59 hallucinations. These results suggest that the specialized instructions are helping to keep the AI's outputs grounded in fact.
What the data reveals
While the legal configuration provides a boost, much of the recent improvement stems from the underlying base model. On Harvey's own LAB benchmark, the move from GPT-6 Astra to GPT-6.2 Astra increased the success rate from 4.6% to 14.6%. The specific legal configuration added an additional 1.6 percentage points to that total.
Julio Pereyra, the head of legal research at Harvey, noted that the model writes more comprehensive and better reasoned briefs. He stated that these briefs are preferred to those of the base model nearly 90% of the time by pairwise judges. This improvement in depth occurred without a regression in accuracy, maintaining the low hallucination rates seen in Astra-class models.
Despite these gains, the model is not yet the leader on every metric. On the Vals AI Legal Research Bench, where practicing lawyers set questions about U.S. law, Astra for Law earned full credit on 53.8% of questions. This is an improvement over its 46% score in September, but it remains a top performer rather than the clear leader on that specific benchmark.
The road ahead for legal AI
OpenAI has framed this release as part of a long-term investment in the legal profession. While the current update focuses heavily on model performance and index size, the company previously highlighted other goals. These included the development of new vendor plugins, community skills, and improving access to justice for the general public.
The company's leadership has emphasized that they are still in the early stages of this journey. Jason Boehmig, who leads OpenAI’s legal products, previously described the launch as just the beginning of their commitment to the field. The current release provides the foundational infrastructure—the model, the instructions, and the data—needed for more complex applications.
As the legal industry continues to adopt frontier models, the focus will likely shift toward how these tools can be integrated into daily workflows. For now, the goal is to provide a reliable, high-performing configuration that can handle complex drafting and research tasks. OpenAI maintains that they are eager to see how the profession will eventually utilize these capabilities.
What to watch
- How the 34% increase in the search index affects research accuracy in non-U.S. jurisdictions.
- Whether OpenAI will introduce new community plugins or custom skills for legal work in the next update.
Try it
If you have a ChatGPT Enterprise account, check your admin console to see if you can request access to the Astra for Law configuration.
Sources and links
- Officialreleased an updated version of Astra for Law openai.com
- Reportinglawnext.com: the article we sourced
First reported by lawnext.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.