OpenAI opens specialized legal AI model to all enterprise users
Astra for Law, a configuration of GPT-6.2, now allows enterprise customers to access a model that leads independent benchmarks for legal reasoning and research.
The short version
- Astra for Law is a configuration of GPT-6.2 using legal-specific instructions and a dedicated search index.
- The model earned full credit on 20.8% of assignments in the Harvey LAB-AA benchmark, beating its closest competitor.
- It averaged 0.03 material hallucinations per assignment, significantly lower than standard models like Claude Opus 5.5.
- The model's search index now covers over 310 million URLs, including 99.9% of published U.S. precedential case law.
Why it matters
For legal professionals, this move lowers the barrier to using high-performing AI that is specifically tuned to minimize 'hallucinations'—the fabrication of facts. It also offers a more cost-effective way to handle high-volume document review by producing shorter, more precise outputs.
Expanding access to specialized AI
OpenAI has announced that every ChatGPT Enterprise customer can now request access to Astra for Law through its Trusted Access Program. This includes the ability for teams to integrate the model into their own internal research and drafting tools via an API. The move marks a shift from limited pilot programs to a broader availability for corporate legal departments.
Astra for Law is not a brand-new foundation model built from scratch. Instead, it is a specialized configuration of OpenAI's GPT-6.2 model. The company wrapped the core model in legal-specific instructions and paired it with a dedicated search index to improve its performance on complex legal tasks.
By making the model available through the standard ChatGPT Enterprise admin console, OpenAI is lowering the barrier for procurement. Legal teams no longer need to navigate bespoke pilot programs to use a tool designed specifically for their unique professional requirements.
Performance on legal benchmarks
The company released these updates following the model's performance in the Harvey LAB-AA benchmark. This test involved 120 assignments across 24 different areas of law, such as reviewing commercial leases and mapping ownership changes in acquisitions. Astra for Law took the top spot, securing full marks on 20.8% of the total tasks.
The results showed that the legal configuration outperformed the plain GPT-6.2 Astra, which only earned full credit on 15.0% of the tasks. It also beat the top non-OpenAI entrant, Muse Wonder Full, which scored 13.1%. OpenAI noted that the legal version cleared the all-pass bar nearly 40% more often than the standard GPT-6.2 Astra.
Beyond just accuracy, the model demonstrated efficiency in its output. Astra for Law averaged roughly 64,000 output tokens per assignment, which is 54% fewer than Muse Wonder Full's 139,000 tokens. This efficiency is significant for legal teams that are often billed based on their total usage volume.
Reducing hallucinations in law
A major hurdle for AI in the legal field is the risk of hallucinations, where a model might invent a fake citation that could damage a lawyer's standing in court. Astra for Law addressed this by averaging just 0.03 material hallucinations per assignment in the LAB-AA benchmark. This rate was tied for the lowest in the comparison and was about half the rate of the regular GPT-6.2 Astra.
The model's performance on reliability was notably higher than other major competitors. For example, Claude Opus 5.5 averaged 0.59 hallucinations per assignment, which is roughly 20 times the rate recorded for Astra for Law. OpenAI stated that the legal model logged no material hallucinations on any tasks that met the all-pass criteria.
These improvements are attributed to a dedicated research team and a growing database of information. Since the system began rolling out to a small group of firms in September 2026, the index has expanded by 34% and now covers over 310 million URLs. This includes more than 99.9% of published U.S. precedential case law and coverage of international jurisdictions like the UK, Australia, and Canada.
Practical use cases for firms
Early users, including firms like Cooley and Sullivan & Cromwell, have already begun deploying the system in various capacities. Lawyers are using the model as a sparring partner to stress-test their legal arguments before they are finalized. It is also being used to separate specific recommendations from requirements in draft legislation.
Another common use case involves converting large stacks of regulatory policy into internal workflows. The model's ability to handle research depth and sensitivity to authority has been highlighted by partners at these firms. They noted that the model demonstrates impressive depth across both litigation and transactional matters.
The broader goal for OpenAI is to provide tools that are built specifically for the work of lawyers who underpin commerce and civil rights. Rather than offering a general-purpose model that must be adapted after the fact, the company is prioritizing legal intelligence as a core research area. This approach aims to build the trust necessary for wider adoption in high-stakes environments like courtrooms and deal rooms.
What to watch
- Whether the lower hallucination rates translate to increased trust in high-stakes courtroom filings.
- How other frontier models like Claude Opus 5.5 and Gemini 4 Argon respond to these new benchmarks.
Try it
If you have a ChatGPT Enterprise account, check your admin console to see if you can request access to Astra for Law through the Trusted Access Program.
Sources and links
First reported by BigGo Finance. 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.