Legal teams using ChatGPT Enterprise can now request access to Astra for Law and connect to the configuration through the API. OpenAI announced the update on November 17, 2026: the model configuration is built on GPT‑6.2 Astra and supplemented with a specialized legal-materials index and instructions for legal research.

The practical reason to pay attention to the news is the combination of a new access channel and published test results. The company is opening the program to all Enterprise customers, whereas, it says, access was previously limited to a small group of law firms. Access requests go through the Trusted Access Program; teams that are already connected can select the identifier gpt-6.2-astra-law. The updated Astra for Law announcement also invites organizations to request API access.
What has changed for enterprise customers
Astra for Law is a specialized configuration of a base model, not a standalone public legal service for all ChatGPT users. According to OpenAI, it combines GPT‑6.2 Astra with an expanded legal search index and settings for finding answers, comparing legal grounds, and drafting text.
The new offering is aimed at ChatGPT Enterprise customers. For a legal team, this means it can test the configuration on its own tasks through a controlled-access program; the API option allows a team to request integration into its tools. The announcement does not promise automatic access for every customer: access must be requested. OpenAI does not specify when requests will be approved.
The company says that since the limited launch in September, it has expanded the index by more than a third, adding over 80 million URLs. According to OpenAI, the index includes more than 310 million URLs and covers over 99.9% of published US case law. These figures describe OpenAI’s index; in practice, legal teams need to check which sources cover their jurisdictions, industries, and document types.
What the published tests show
OpenAI reports results from two evaluations that test different sets of legal tasks. LAB-AA tests 120 tasks across 24 areas of law. Examples include preparing deposition questions, reviewing commercial contracts, and analyzing ownership-change options in an acquisition. A perfect score requires meeting every task requirement without making material claims unsupported by the source documents.
In this test, Astra for Law earned a perfect score on 20.8% of tasks, GPT‑6.2 Astra on 15.0%, and Muse Wonder Full on 13.1%. This result shows the share of tasks that received a perfect score under these specific conditions. It helps compare models on the defined set of legal assignments, but teams will need their own documents and review criteria to evaluate their workflows.
The second test, Vals AI’s Legal Research Bench, evaluates questions about US federal and state law. OpenAI reports that in the November 16 evaluation, the configuration earned a perfect score on 112 of 208 questions, or 53.8%; in September, the figure was about 46%. The conditions and question set differ from LAB-AA, so these percentages describe a separate evaluation, not the model’s universal accuracy.
The company also says that in LAB-AA, Astra for Law generated an average of around 64,000 output tokens per task, compared with 139,000 for Muse Wonder Full. For teams paying based on usage volume, this could be an important pilot metric alongside answer quality. Actual costs depend on the specific integration and access terms.
How to apply the results in a legal team
A perfect score requires meeting every requirement of a specific task. During a pilot, it is therefore useful to track not just an overall impression of an answer, but also which requirements it met, what sources it relied on, and where a specialist had to correct or supplement the result. This lets a team compare its results with the type of evaluation described in LAB-AA.
For legal research, teams should separately check that sources are up to date and relevant to the required jurisdiction. OpenAI says it has expanded the index to include materials from the United Kingdom, Australia, and Canada, as well as coverage of US case law. A list of countries alone does not indicate the depth of coverage for a specific question; during an operational evaluation, teams can compare the model’s citations with primary legal documents.
The main change is that all ChatGPT Enterprise customers can now request access to the specialized legal configuration, including teams that want to integrate it through the API. The published tests provide measurable reference points for selecting pilot tasks: US legal research, contract analysis, or preparing case materials. They also suggest which metrics to check together: requirement completion, support from citations, and the volume of generated text.