Tools

AI-generated text

OpenAI Expands Access to Its Astra for Law Configuration

OpenAI has updated Astra for Law and opened access requests to all ChatGPT Enterprise customers, including for use through the API.

OpenAI Expands Access to Its Astra for Law Configuration

OpenAI has updated its Astra for Law legal model configuration and enabled all ChatGPT Enterprise customers to request access through the Trusted Access Program. Access can be requested through the admin console and for API use.

Astra for Law is built on GPT‑6.2 Astra. The configuration combines the model with an expanded legal search index and instructions for researching legal questions, weighing authorities and writing answers. OpenAI says the update improves legal reasoning and search.

The company initially introduced the configuration in September to a small group of law firms, including Cooley, Ropes & Gray and Sullivan & Cromwell. Since then, the search index has grown by 34%, adding more than 80 million URLs, including sources from jurisdictions such as the United Kingdom, Australia and Canada. It now contains more than 310 million URLs and, according to OpenAI, covers more than 99.9% of published precedential U.S. case law.

Results on complex legal assignments

LAB-AA, Artificial Analysis’s version of Harvey’s Legal Agent Benchmark, assesses 120 assignments across 24 areas of law. They include drafting deposition outlines, reviewing commercial leases and planning ownership changes in acquisitions. To receive full credit, an answer must satisfy every requirement without making a material claim unsupported by the source documents.

Astra for Law earned full credit on 20.8% of assignments, ahead of GPT‑6.2 Astra at 15.0% and Muse Wonder Full at 13.1%. That meant nearly 40% more fully credited assignments than GPT‑6.2 Astra and about 59% more than Muse Wonder Full, the highest-scoring non-OpenAI model.

Astra for Law used an average of about 64,000 output tokens per assignment, compared with 139,000 for Muse Wonder Full. The 54% reduction can matter for costs under usage-based pricing. Output tokens include text generated as a model reasons and writes.

In the Artificial Analysis comparison, Astra for Law and GPT‑6 tied for the lowest rate of material hallucinations, averaging 0.03 per assignment. Regular GPT‑6.2 Astra averaged 0.06, while Claude Opus 5.5 averaged 0.59. Astra for Law also averaged 0.06 minor hallucinations per assignment, compared with 0.51 for Claude Opus 5.5. OpenAI says Astra for Law produced no material hallucinations on assignments that received full credit.

Brief writing and legal research

LAB: USCL, co-developed by Harvey and Grubel, tests how models research and write a brief for a real U.S. federal case. In one version, the model must find relevant decisions itself; in another, it receives the case record and precedents cited by the original lawyer. AI graders compare the briefs using criteria reviewed and calibrated by lawyers.

Astra for Law was among the top performers in both versions. When conducting its own research, it scored 1,844 points, compared with 1,817 for Opus 5.5. When given the precedents, their results were effectively even, at 1,883 and 1,882 respectively.

Vals AI’s Legal Research Bench uses questions about U.S. federal and state law developed by practicing lawyers. Astra for Law received full credit on about 46% of questions in September; in the November 16 evaluation, it did so on 112 of 208 questions, or 53.8%. That was an increase of about eight percentage points in less than two months. Including partial credit, the configuration earned 90.9% of the available points.

On Harvey’s own Legal Agent Benchmark, which is separate from LAB-AA, GPT‑6 Astra met every requirement on 4.6% of assignments, GPT‑6.2 Astra on 14.6%, and the legal configuration on 16.2%.

Uses and access

OpenAI says early users at law firms and within its own legal team are testing the tool. Uses include finding counterarguments to legal positions, interpreting obligations in proposed legislation and mapping rules that apply to new business activities. It has also been used to turn regulatory procedures and company policies into a workflow, and to build an annotated guide to legislation with explanations and references.

Customers who already have access can select the gpt-6.2-astra-law configuration in ChatGPT. API access will allow customers to incorporate the configuration into their own tools.