Astra for Law hits the mark and arrives at law firms

Astra for Law is OpenAI's new platform for legal work, launched on Thursday, the 17th. The infrastructure combines GPT 6 Astra, released in early September, with tools and context tailored exclusively for legal professionals. The initial audience includes law firms and legal tech companies.

The proposal draws attention. However, the accuracy numbers call for careful reading before any enthusiasm.

Astra for Law brings its own index with 230 million pages

The technical highlight is called the Legal Search Index. It is a proprietary search tool focused on legal databases in the United States.

In addition, the index integrates with the Free Law Project repository. This project offers free access to primary legal materials.

Coverage exceeds 230 million web pages. Therefore, the model queries a curated database, rather than relying solely on common search.

Notice the architectural decision. Retrieval from a specialized source tends to reduce hallucinated answers in high-precision domains.

Zero Data Retention becomes a selling point

Privacy appears as the second pillar. The platform runs in an environment with Zero Data Retention, known by the acronym ZDR.

In practice, prompts, submitted data, and responses stay off the company's servers. Likewise, this content stays out of the training of new models.

This point carries a lot of weight in the industry. After all, attorney-client privilege blocks the adoption of any tool that stores conversations.

Astra for Law arrives with 26 plugins and Word integration

The package includes 26 new plugins for the legal context. Relativity and Clio are among the examples cited.

At the same time, ChatGPT's integration with Word has been rolled out to everyone. As a result, the AI now works directly on documents within Microsoft's editor.

This combination explains the strategy. The model steps in where the lawyer already works, with less switching between tools.

The numbers show progress, and also the current limit

Here is the part that deserves more attention. On the Vals AI Legal Research Bench, the platform passed the correctness validation on 54% of the questions.

For comparison, GPT 6 Astra with common web search scored 38.7%. In other words, the gain is real and significant.

In addition, the tool correctly extracted 54% of the most relevant passages from the court decisions in the audited set.

Even so, the result remains far from the reliability required in day-to-day legal work. Nearly half of the questions still come out with a problem.

What legal tech teams should consider before integrating

First, treat the output as a verifiable draft. Every citation needs a link to the primary source and human review.

Then, build verification into the product itself. An automatic step that checks whether the cited decision exists significantly reduces risk.

Also consider the geographic scope. The index covers United States databases, and use in Brazil requires a local source of case law.

In addition, carefully read the ZDR terms. Contract, audit log, and data jurisdiction need to be clear for the end client.

Finally, measure with your own cases. A set of real questions from your firm is worth more than any public benchmark.

Limited availability for now

Current access is through the Trusted Access program, in ChatGPT and Codex. Only selected firms take part.

The model appears in the selector as GPT 6 Astra Law. API access, meanwhile, comes later, with no date announced yet.

What to watch in the coming months

Keep an eye on the API release and the pricing model. This detail determines viability for legal tech products.

Also follow independent evaluations with databases from other countries. In the meantime, it's worth prototyping workflows with mandatory human review from day one.

Follow our profile on Instagram!