OpenAI launches Astra for Law for AI-powered legal workflows
OpenAI has launched Astra for Law, a legal-specific system built around GPT-6 Astra that combines frontier-model reasoning with a dedicated U.S. legal search index covering more than 230 million URLs.
In OpenAI’s 200-question Legal Research Bench, Astra for Law achieved 54.0% overall correctness versus 38.7% for GPT-6 Astra with standard web search, an absolute increase of 15.3 percentage points.
Data Studios calculates the relative improvement at 39.5%, using (54.0 − 38.7) / 38.7.
The comparison isolates an important part of the product design: Astra for Law does not depend on a new legal foundation model alone, but combines GPT-6 Astra with specialized retrieval, legal instructions, integrations and controls intended for professional legal workflows.
The benchmark remains a vendor-reported evaluation rather than an independent measure of performance across legal practice, and its 54.0% absolute score also establishes a significant limitation: specialized retrieval improves the reported result without making professional verification optional.
··········
ASTRA FOR LAW COMBINES GPT-6 ASTRA WITH SPECIALIZED LEGAL RETRIEVAL.
The system changes the information supplied to the model before legal reasoning begins, using a dedicated index rather than relying only on ordinary web search.
OpenAI says the legal search index covers more than 230 million URLs across U.S. case law, statutes, regulations, court rules and administrative decisions.
Material from CourtListener and the Free Law Project is incorporated into the retrieval environment, giving the system access to a corpus specifically structured around legal authorities.
The resulting workflow can be represented as legal question → legal search → authority retrieval → relevant-passage retrieval → GPT-6 Astra reasoning → supported output → professional review.
A general web search can surface primary authorities alongside law-firm commentary, summaries, news coverage and other secondary material.
A dedicated legal retrieval layer instead narrows the information environment toward sources that can establish legal authority before GPT-6 Astra performs synthesis and analysis.
This architecture separates two sources of error that are often conflated in discussions of legal AI: the model can reason poorly from good evidence, but it can also reason competently from an incomplete or incorrectly retrieved set of authorities.
........
Metric | Astra for Law | GPT-6 Astra + web search | Difference |
Overall correctness | 54.0% | 38.7% | +15.3 pp |
Relative improvement | — | — | +39.5% |
Retrieval environment | Specialized U.S. legal index | General web search | Domain-specific retrieval |
Indexed corpus | 230M+ URLs | Not directly comparable | Legal-source corpus |
Legal-specific configuration | Yes | No | Specialized instructions and workflow |
........
The 39.5% figure is a Data Studios calculation, while the underlying 54.0% and 38.7% scores are reported by OpenAI.
It should not be interpreted as a 39.5% improvement across every legal task, jurisdiction or professional workflow because the comparison applies specifically to OpenAI’s 200-question Legal Research Bench.
··········
RETRIEVAL PERFORMANCE EXPLAINS PART OF THE BENCHMARK DIFFERENCE.
OpenAI reports improvements at both authority discovery and passage retrieval, two stages that occur before the model produces its final legal analysis.
For questions focused on case law, OpenAI reports that Astra for Law retrieved 24% more reference cases than the comparison system.
The company also reports up to 54% more relevant passages from the correct decisions.
Those measurements describe separate operations and therefore cannot be added into a combined 78% improvement.
Authority discovery determines whether the appropriate case enters the model’s working evidence set, while passage retrieval determines whether the relevant language inside that decision is actually supplied to the reasoning process.
A failure at either stage can degrade the final answer without necessarily demonstrating a failure of the foundation model itself.
This provides a more precise interpretation of Astra for Law’s architecture: some of the reported performance gain comes from changing the evidence available to GPT-6 Astra, rather than simply asking the same model to reason harder over the same search results.
Corpus size alone does not establish retrieval quality. A 230-million-URL index still depends on coverage, freshness, ranking, jurisdictional filtering, citation integrity and the system’s ability to identify the legally relevant portion of a document.
··········
ASTRA FOR LAW EXTENDS FROM RESEARCH INTO LEGAL WORKFLOWS.
OpenAI is building a professional application layer around retrieval and reasoning, including integrations that can connect the model to existing legal software and matter data.
Astra for Law is intended for law firms as well as legal-technology companies, allowing the underlying capabilities to appear either directly in OpenAI workflows or inside third-party products.
OpenAI identifies Harvey and Legora among legal AI companies working with the technology through its API ecosystem.
The company also reports 26 new legal plugins, including integrations involving platforms such as Relativity and Clio.
The operational distinction is significant because professional legal AI increasingly depends on what information a system can retrieve from an organization’s existing environment, not merely on what a user manually places into a prompt.
An integrated workflow can potentially retrieve matter context, locate legal authorities, perform a defined analysis and return the result to the professional environment in which the work is being conducted.
........
System layer | Function | Principal constraint |
Foundation model | GPT-6 Astra reasoning | Reasoning errors remain possible |
Legal retrieval | 230M+ URL U.S. legal index | Coverage and retrieval quality |
Legal configuration | Specialized analysis and writing instructions | Configuration does not guarantee correctness |
Integrations | 26 legal plugins | Permissions and available connector data |
Application layer | OpenAI and third-party legal workflows | Implementation varies by product |
Professional control | Lawyer review and source verification | Human verification remains necessary |
........
Integrations also increase the importance of permission design.
Legal matters can contain privileged communications, personal information, confidential commercial documents and material subject to contractual or regulatory restrictions.
OpenAI says Astra for Law includes controls intended for confidential client work, but an actual deployment still has to determine which data the system can access, who can authorize that access, whether information can cross matter boundaries, how information is retained and what external integrations receive it.
A system that can automatically obtain richer matter context can reduce manual work while simultaneously increasing the consequences of an incorrectly configured permission.
··········
THE 54% ABSOLUTE SCORE DEFINES THE CURRENT PROFESSIONAL BOUNDARY.
The benchmark supports specialized retrieval as a performance improvement, but it does not support treating Astra for Law as autonomous legal research.
Astra for Law’s 54.0% overall correctness means that the system did not receive full credit on 46.0% of the questions in OpenAI’s reported evaluation.
The 46.0% figure is a Data Studios calculation from 100% − 54.0%; it does not mean that 46% of ordinary Astra for Law outputs will be wrong, because benchmark composition, scoring and difficulty cannot be mechanically generalized to real legal matters.
The legal index is also described around U.S. legal materials, so equivalent coverage or performance should not be assumed for other jurisdictions without separate evidence.
Retrieving an authentic judicial decision does not establish that the decision remains good law, controls in the relevant jurisdiction or supports the proposition for which the system uses it.
Likewise, a technically accurate summary may still be insufficient when an issue depends on procedural posture, factual distinctions, local rules or subsequent authorities.
The practical decision rule is therefore narrower than the headline benchmark improvement suggests: Astra for Law can be used to accelerate retrieval and analysis when its authorities remain traceable and professionally reviewable, but the reported evidence does not justify replacing source verification or legal judgment.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
