Today, we’re releasing an updated Astra for Law that is state-of-the-art on Harvey LAB-AA and a top performer on LAB: USCL (co-developed by Harvey(opens in a new window) and Grubel(opens in a new window)) and Vals AI’s Legal Research Bench. We are also enabling all ChatGPT Enterprise customers to request access through the Trusted Access Program, including through the API.
Astra for Law is a model configuration of GPT‑6.2 Astra, our latest, most powerful frontier model. It pairs GPT‑6.2 Astra with an expanded legal search index and instructions for researching a question, weighing the authorities, and writing the answer. Today’s configuration earned full credit on the LAB-AA tests nearly 40% more often than did the runner-up, GPT‑6.2 Astra.
Since we introduced Astra for Law in September to a small group of firms, including Cooley, Ropes & Gray and Sullivan & Cromwell, our research team has made rapid innovations to improve its legal reasoning and expand its legal search index by over a third. On Vals AI’s Legal Research Bench(opens in a new window), the share of research answers that passed every check rose from about 46% to 53.8% in less than two months.
As the world grows more complex, lawyers occupy an increasingly important role. They shape nearly every system we rely on: they negotiate the agreements that move commerce, defend people’s rights in courts, and draft the laws we all live by. That’s why we’ve made law a research priority at OpenAI: to put the best intelligence in the hands of the best lawyers.
Harvey’s Legal Agent Benchmark (LAB) tests whether a model can take a legal assignment from start to finish. LAB-AA(opens in a new window) is Artificial Analysis’s version of that test. It puts the models through a ringer of 120 assignments developed by Harvey across 24 areas of law – everything from drafting deposition outlines to reviewing commercial leases to strategizing ownership changes in acquisitions. To earn full credit, an answer must meet every requirement of an assignment without hallucinating, i.e. making a material claim that is not supported by the source documents.
Astra for Law ranked first, earning full credit on 20.8% of assignments, ahead of GPT‑6.2 Astra at 15.0% and Muse Wonder Full at 13.1%.
Astra for Law ranked first at 20.8%, earning full credit about 59% more often than Muse Wonder Full, the highest-scoring non-OpenAI model.
Astra for Law also earned that top score with 54% less output per assignment than the highest-scoring non-OpenAI model in the evaluation, averaging about 64,000 output tokens per assignment, compared with 139,000 for Muse Wonder Full. For legal teams paying by usage, this can mean the difference between staying within budget and blowing through it.

Each dot is a model configuration. Higher means more assignments earned full credit; farther left means fewer output tokens, which include the text a model generates as it reasons and writes. Astra for Law averaged about 64,000 output tokens per assignment and earned full credit on 20.8%, compared with 139,000 and 13.1% for Muse Wonder Full. No configuration shown matched Astra for Law’s score using fewer output tokens.
Hallucinations are unacceptable in legal work. A single invented citation can cost a lawyer a client’s trust, credibility in court, and a reputation built over years. That’s why reducing hallucinations has been a priority for Astra for Law. In Artificial Analysis’s latest LAB-AA comparison, Astra for Law tied (with GPT‑6) for the lowest rate of material hallucinations: 0.03 per assignment, compared with 0.59 for Claude Opus 5.5 (the next lowest non-OpenAI model). Notably, Astra for Law’s performance on this eval included zero material hallucinations for tasks that met the all-pass criteria.

Astra for Law averaged 0.03 material and 0.06 minor hallucinations per assignment. Claude Opus 5.5 averaged 0.59 material and 0.51 minor hallucinations per assignment, about 20 times as many material hallucinations and 8.5 times as many minor ones.
Across the same 120 assignments, Artificial Analysis counted hallucinated claims that could mislead a reader on a substantive point. Astra for Law averaged 0.03 such errors per assignment, tied for the lowest rate in the comparison and roughly half the rate for regular GPT‑6.2 Astra (0.06).
Every OpenAI configuration tested had fewer hallucinations than Claude Opus 5.5, the best non-OpenAI result on this measure. Opus 5.5 averaged 0.59 material hallucinations per assignment, about 20 times as many as Astra for Law.
Brief writing is one of the most complex long-horizon processes in law, requiring clear reasoning, research, and writing to come together to produce an excellent argument.
Harvey and Grubel(opens in a new window) co-developed LAB: USCL(opens in a new window) to test whether a model can research and write a brief in a real U.S. federal case. Each model steps in just before a brief was filed and writes for the same party on the same issue. In one version, the model has to find the relevant cases itself. In the other, it gets the record and the precedents the original lawyer cited. AI graders compare the briefs head to head using criteria lawyers reviewed and calibrated. Astra for Law earned one of the top scores in both versions.

Farther right means the briefs earned higher ratings. When the models researched the law themselves, Astra for Law scored 1,844 to Opus 5.5’s 1,817. Given the precedents, they were effectively even at 1,883 and 1,882.
The first result (“Research and draft,” on the left) matters because better legal research is one of the core improvements in this release. We focused on improving this result because a litigator rarely starts with a list of the right cases. Here, each model had to work out which decisions mattered and use them to write an argument the court could follow. On that assignment, Harvey’s graders preferred the briefs from today’s Astra for Law to those from other leading frontier models, such as Opus 5.5 and Gemini 4 Argon. We can see more noticeable improvements in legal research in the Vals AI evaluation, discussed next.
We’ve made law a research priority at OpenAI, dedicating a team of researchers to accelerate innovation for frontier legal intelligence. One early measure of their progress comes from Vals AI’s Legal Research Bench(opens in a new window), where practicing lawyers set questions about U.S. federal and state law. An answer has to cover every required point and pass the source checks. In September, Astra for Law earned full credit on about 46% of questions. In Vals AI’s November 16 evaluation, it earned full credit on 112 of 208 questions (53.8%), a gain of about eight percentage points in under two months. Across all questions, it earned 90.9% of the available points, including partial credit.

In less than two months, the share of legal research questions on which Astra for Law earned full credit rose from about 46% to 53.8%, a gain of about eight percentage points.

Astra for Law (green) sits in the small cluster of top performers on Vals AI’s Legal Research Bench.
Building Astra for Law as a model configuration lets us stay on the bleeding edge of the frontier, in this case GPT‑6.2 Astra. On top of this foundation, we iteratively refine Astra for Law’s specific settings and instructions in response to practitioner feedback, teaching it how to reason and write like a seasoned lawyer. We also continuously hone its tools, such as the legal search index. Since September, we’ve grown the index by 34%, adding over 80 million new URLs, including across international jurisdictions such as the UK, Australia and Canada. Today, the updated index spans over 310 million URLs and covers more than 99.9% of published U.S. precedential case law(opens in a new window).
The below chart shows how the model configuration takes advantage of both the speed of frontier innovation and builds on top of that the specific requirements that make it powerful for lawyers. On Harvey’s own Legal Agent Benchmark, GPT‑6 Astra met every requirement on 4.6% of assignments. GPT‑6.2 Astra tripled performance, completing 14.6% under the same conditions. With the legal configuration, that rose to 16.2%, completing 11% more questions than GPT‑6.2 Astra.

This chart shows both the rate of development in frontier intelligence, as well as the impact of the model configuration. The full-credit rate more than tripled from GPT‑6 Astra to GPT‑6.2 Astra, then rose another 11% with the legal configuration. This test comes from Harvey’s own LAB (separate from the previously discussed LAB-AA).
The most interesting possibilities for Astra for Law are being shaped by the lawyers who are using it for their day-to-day work. Across law firm customers and OpenAI’s own legal team, early users are building references around their practice, inviting tougher challenges to their arguments, and finding new ways to respond to changing legislation.
For example, early users at OpenAI are using Astra for Law to:
-
Give your argument its strongest opponent. A common use for Astra for Law is as a sparring partner. A competition lawyer gives it a reading of a statute and asks it to argue the other side, drawing on legislative history, court decisions and versions of the law in other languages.
-
Find the line between a recommendation and a requirement. A policy lawyer looking at a proposed oversight bill wanted to know how much power an independent reviewer would have. If the reviewer called for a fix, would the developer have to make it? For questions like these, the lawyer puts a plain-English account of each obligation beside the bill’s exact words, so she can make fast judgment calls about which reading the text supports.
-
Start the first call with outside counsel on the hardest question. When the business is considering entering an unfamiliar industry, a product lawyer uses Astra for Law to map which rules would apply to the proposed activity. Instead of spending the first call asking for high-level advice, the lawyer can bring outside counsel what’s actually at issue.
-
Turn a stack of policies into a workflow. A legal program manager used Astra for Law to turn regulatory procedures and company policy into a workflow for colleagues handling incoming cases, enabling them to automate when to ask for more information, when to close the case and when to escalate it to a lawyer.
-
Build the guide you wish came with the statute. A competition lawyer paired Astra for Law with ChatGPT’s site-building tools to annotate legislation the team uses regularly. A colleague reading a provision can hover over a term for its definition, follow a link to the materials that explain it and see what the team advised the last time that provision came up.
Starting today, every ChatGPT Enterprise customer can request Astra for Law through the Trusted Access Program in the admin console. If your team already has access, select gpt-6.2-astra-law in ChatGPT. Customers can also request API access to build Astra for Law into your own tools using this form.
The legal profession is at an inflection point. As technology changes what businesses can do and the world becomes increasingly complex and interconnected, lawyers have an unprecedented opportunity to shape how we make sense of the moment, not just for the businesses they advise but for all of us who live by the rules they help write.
We believe the best lawyers, equipped with the best intelligence, can meet that moment. We’re only at the beginning. We can’t wait to see where you take the profession.


