AI News
AI News AgentModel releaseOpenAI4 min read

OpenAI improves Astra for Law for legal research

OpenAI has updated Astra for Law, a GPT-6.2 Astra configuration focused on legal research and drafting. The system leads several legal evaluations, reduces hallucinations, and can now be requested by all ChatGPT Enterprise customers, including through API access.

OpenAI has updated Astra for Law, a GPT-6.2 Astra configuration designed to research legal questions, evaluate sources, and draft responses. From now on, all ChatGPT Enterprise customers can request access, including through the API.

The update delivered strong results in three legal-work evaluations. According to tests cited by OpenAI, the system performs well on tasks ranging from reviewing contracts and preparing depositions to researching precedents and drafting arguments for a federal case in the United States.

What it achieved in testing

In LAB-AA, an evaluation of 120 legal assignments created by Harvey and adapted by Artificial Analysis, Astra for Law ranked first. The test covers 24 areas of law and requires the system to meet every requirement without inventing claims unsupported by the documents.

  • Astra for Law received a perfect score on 20.8% of assignments.
  • GPT-6.2 Astra did so on 15%.
  • The next non-OpenAI model reached 13.1%.

The legal configuration also used less text to complete each task: around 64,000 output tokens per assignment, compared with 139,000 for the highest-scoring non-OpenAI model. A token is a unit of text used to calculate model usage. For a team that pays based on usage, producing less text can reduce costs.

Accuracy is especially important in law. An invented citation or unsupported interpretation can cause a lawyer to lose time, credibility, or even a case.

In the same evaluation, Astra for Law recorded an average of 0.03 material hallucinations per assignment, tying with GPT-6 for the lowest result in the comparison. OpenAI defines these hallucinations as invented claims that could mislead someone about a substantive point. Claude Opus 5.5, the best-performing non-OpenAI model on this metric, recorded 0.59.

That does not mean the system can work without supervision. The figures come from specific evaluations and do not guarantee that a real-world response will be correct in every case.

Improvement in legal research

In LAB: USCL, a test developed by Harvey and Grubel, models must research and draft a brief for one party in a real federal case in the United States. In one version, they must find the relevant precedents themselves instead of receiving a prepared list.

Evaluators preferred Astra for Law's brief over those produced by other advanced models, including Claude Opus 5.5 and Gemini 4 Argon, in the version that combines research and drafting. The test aims to measure something more demanding than answering questions: building a complete legal argument from relevant sources.

In Vals AI's Legal Research Bench, created with questions from practicing lawyers about US federal and state law, the model received a perfect score on 112 of 208 questions, or 53.8%, in the November 16 evaluation. In September, it had reached approximately 46%.

OpenAI also expanded the legal search index associated with the system. Since September, it has added more than 80 million URLs, a 34% increase, bringing the total to more than 310 million. The index includes sources from the United States, United Kingdom, Australia, and Canada, and covers more than 99.9% of published US case law that sets precedent, according to the company.

What a legal team can use it for

The idea is not to replace professional judgment, but to speed up tasks that usually consume many hours. OpenAI describes uses including:

  • Arguing the opposing position to stress-test a strategy.
  • Comparing the exact wording of a rule with its practical obligations.
  • Identifying which rules could affect a new business activity.
  • Turning policies and procedures into workflows for deciding when to request information, close a case, or escalate it to a lawyer.
  • Creating internal guides that explain a law and link each interpretation to its sources.

For end users, the most important change is indirect: if these tools perform as the tests indicate, lawyers can reach relevant precedents sooner, prepare stronger arguments, and spend less time on mechanical tasks. But every conclusion will still require human review, especially when rights, contracts, or decisions with legal consequences are involved.

ChatGPT Enterprise customers can request access from the admin console through the Trusted Access Program. If they already have authorization, they can select gpt-6.2-astra-law in ChatGPT and request API access to integrate it into their own tools.

What to watch now is whether the performance seen in controlled tests holds up in real cases, with different jurisdictions, incomplete documents, and ambiguous questions. In legal work, getting more answers right matters, but so does knowing when an answer is not reliable enough to support a decision.