GPT-6 Astra connects directly to U.S. law

OpenAI is launching Astra for Law, a GPT-6 Astra configuration paired with an index of 230 million U.S. legal sources and tools designed for law firms.

A persuasive legal answer is worthless if it relies on a nonexistent ruling, overturned precedent, or outdated law. With Astra for Law, OpenAI is attempting to bring generated analysis closer to the authoritative sources needed to support it. The service combines GPT-6 Astra with instructions tailored to legal analysis and writing, settings designed for in-depth work, and a new search index. It will initially be offered to selected law firms through ChatGPT and Codex, with API access planned for later.

Astra for Law is not a separate foundation model trained specifically on legal data. It is an OpenAI-maintained configuration built around GPT-6 Astra. The company combines the general-purpose model with a specialized working method, legal sources, and connections to software already used by legal professionals.

That distinction matters. Its legal behavior does not come solely from knowledge memorized during training. It also depends on the instructions governing its response, the requested reasoning level, the documents retrieved at query time, and the firm’s own rules.

The main addition is the Legal Search Index. It can search U.S. case law, statutes, regulations, court rules, and administrative decisions. OpenAI says the corpus covers more than 230 million URLs, with new sources added daily.

Part of the case law comes from CourtListener, a service operated by the nonprofit Free Law Project. That collection claims to cover more than 99.9% of published U.S. precedential opinions. It gives Astra for Law access to a public resource that is far more structured than an ordinary web search.

The figure of 230 million, however, does not represent 230 million unique rulings. It combines multiple categories of sources and counts URLs rather than deduplicated legal documents. The number indicates the index’s breadth, but not the quality of its rankings, the completeness of each jurisdiction, or the update frequency of every source.

Its coverage is also limited to the United States. The announcement does not describe an equivalent index for French, European, British, or other legal systems. An international firm may still use GPT-6 Astra’s general capabilities for those areas, but it will not necessarily receive the same document-retrieval infrastructure.

OpenAI does not present its index as a complete replacement for Westlaw, LexisNexis, CoCounsel, or firms’ internal knowledge bases. The company says it complements the licensed content and specialized products to which firms already subscribe. Thomson Reuters is itself participating in the launch through several integrations.

Astra for Law can start from the facts of a matter, search for potentially applicable authorities, retrieve relevant passages, and incorporate them into an analysis. The sources remain accessible so lawyers can open the original documents and review the proposed reasoning themselves.

The specialized instructions also direct the system to distinguish a ruling’s central holding from secondary observations, search for material that weakens a position, and flag uncertainty. In transactional work, it can examine the consequences of a contractual exception, compare multiple clauses, or prepare proposed revisions.

The goal is therefore not merely to produce an answer that sounds more legal. OpenAI is trying to structure a process that moves from the question to the sources, then from those sources to a recommendation that can be reviewed.

The company evaluated this configuration on 200 U.S. questions drawn from a private validation set for Vals AI’s Legal Research Bench. The test measures whether a system can retrieve the correct sources and supporting passages, then satisfy several criteria relating to the final answer.

At its maximum reasoning level, Astra for Law satisfied all correctness criteria on 54% of the questions. GPT-6 Astra paired with standard web search reached 38.7% under the same conditions. The difference amounts to 15.3 percentage points, or a relative improvement of approximately 40%.

On questions focused on case law, the new configuration retrieved 24% more reference decisions. In an audited subset, it also extracted up to 54% more relevant passages from the correct judicial opinions.

These improvements primarily measure the contribution of the index and legal configuration compared with GPT-6 Astra using the web. They do not provide a comprehensive ranking against competing legal services. OpenAI does not compare Astra for Law with Westlaw, LexisNexis, CoCounsel, Harvey, or Legora under the same protocol.

The validation set is also private. The questions, expected answers, and detailed failure cases were not released with the announcement. An outside team therefore cannot fully reproduce the results or examine the legal areas in which the system struggles most.

The 54% score also calls for caution. Even with the index and maximum reasoning level, the configuration fails to satisfy every criterion on nearly one question out of two. The evaluation is strict, and a failure may indicate a missing component rather than an entirely incorrect answer, but the result does not support treating Astra for Law as an authority without review.

Vals AI’s public leaderboard provides another point of comparison, with several models scoring near or slightly above 55% on Legal Research Bench. Those figures cannot be directly compared with Astra for Law’s 54%, however. OpenAI used a private portion of the benchmark and evaluated its full configuration, while the public table reflects different runs and testing conditions.

OpenAI provides no figures for latency, average search cost, the number of sources consulted, or response length at the maximum setting. A more thorough analysis may take longer and consume more tokens. These factors will become important when the service reaches the API.

Initial access will be provided through Trusted Access, a program reserved for eligible law firms. Astra for Law will appear in the model selector as “GPT-6 Astra