Astra for Law is the system around the model

The notable part of Astra for Law is not simply a new model in a legal interface. OpenAI has combined GPT-6 Astra with legal-analysis instructions, settings for thorough work, a U.S. legal search index and connected tools for firm workflows.

OpenAI says the index covers more than 230 million URLs across case law, statutes, regulations, court rules and administrative decisions. That creates a different proposition from ordinary web search: a system intended to retrieve authorities and surface relevant passages while lawyers retain responsibility for verification and judgment.

That distinction matters because legal work is shaped by the path to an answer as much as the answer itself. A useful system needs to find relevant authority, expose the supporting passage, work within a firm’s tools and leave room for professional challenge.

OpenAI’s framing is open and composable: firms can use specialist products, bring their own tools and knowledge, and adapt community-built skills. The model sits inside that arrangement; it is not the whole product.

OpenAI says Astra for Law passed the evaluation’s overall correctness check on 54.0% of questions, compared with 38.7% for GPT-6 Astra using web search alone. It describes that difference as a 40% relative improvement.

The stronger reading is also the more cautious one. These are supplier-reported results from a private 200-question Vals AI Legal Research Bench validation set, run at the highest reasoning effort for both systems. They indicate what OpenAI observed in that setup; they do not independently establish legal reliability or professional fitness.

The benchmark is useful as a directional receipt for the product design. It compares the complete Astra for Law configuration with GPT-6 Astra using web search alone, so it points to the value of the surrounding legal system as well as the model.

It should not be stretched beyond that. The evaluation is private, the conditions are specific and the claim comes from OpenAI. In legal work, a stronger retrieval loop still leaves the duty to inspect authorities, test the reasoning and apply professional judgment.

Astra for Law passed the evaluation’s overall correctness check on 54.0% of questions, compared with 38.7% for GPT-6 Astra using web search alone.

OpenAI

The product evidence is clearest when it stays close to the launch materials. OpenAI describes legal-analysis and writing instructions, settings for thorough work, and a legal index that searches U.S. authorities across more than 230 million URLs, with sources added daily.

That is a workflow claim rather than a claim that a model has solved legal work. The intended gain is better access to relevant material and context; the legal conclusion still has to withstand lawyer verification.

OpenAI’s legal search proposition starts with a question that needs to be bounded by jurisdiction, issue and context. The system then draws on legal instructions, settings and indexed authorities to retrieve material that may be relevant.

The key step follows retrieval. We still need to inspect the cited passage, test whether the authority applies and decide what the evidence means for the matter at hand. The workflow can accelerate research; it does not transfer judgment.

For now, Astra for Law is not a broadly available legal product. OpenAI says it will initially be offered to selected law firms through Trusted Access in ChatGPT and Codex, with API availability coming soon.

The ecosystem matters here. OpenAI announced 26 partner-built plugins, 9 community plugins and 47 custom skills. Those numbers describe an expanding set of possible connections, not a guarantee that every workflow is ready or dependable for every firm.

The model matters, but the legal product is the system that helps us find, inspect and verify the authority around it.

AI News analysis of OpenAI’s launch materials

Astra for Law is most interesting as a systems launch. OpenAI is combining model capability with retrieval, instructions, controls and specialist connections in an attempt to make legal AI more useful in the flow of work.

The test is not whether the model can produce a confident answer. It is whether the system helps us locate better evidence, understand its limits and preserve the review discipline that legal work requires. Wider access, independent evaluation and dependable use in real workflows will determine how much of the launch holds up.