OpenAI Astra for Law: the model is not the product, the index is

OpenAI announced Astra for Law on 17 September 2026. In the announcement's own words, as quoted by Artificial Lawyer, it "pairs GPT-6 Astra with instructions for legal analysis and writing, settings for thorough work, and a new Legal Search Index." It goes first to selected US law firms through OpenAI's Trusted Access programme in ChatGPT and Codex, with API access described as coming soon.

A note on sourcing before anything else. OpenAI's announcement page blocks automated readers, so every figure below is taken from the trade coverage that quotes it directly: Artificial Lawyer, Legal IT Insider, MLQ and PYMNTS. Where I quote OpenAI, it is OpenAI's wording as those outlets reproduced it.

I build legal AI, so I read this one more closely than most model launches. The interesting thing about it is not the model. It is what OpenAI chose to put around the model, and what that says about where the value in legal AI actually sits.

A configuration, not a new model#

The most useful sentence in the coverage comes from MLQ, which describes Astra for Law as "a product configuration and workflow foundation, not a separately trained base model."

That is worth taking seriously. There are three components: the GPT-6 Astra model that anyone with access already has, a set of instructions for legal analysis and writing, and a search index built for law. Two of the three are not model work at all. One is prompt and settings engineering. The other is a corpus.

When a lab with OpenAI's resources decides that the way to make its best model good at law is to give it a better library and better instructions, rather than to train a legal model, that is a statement about where the leverage is. I made the same argument from the other end in the architecture behind a source-grounded legal AI assistant: corpus preparation beats model selection. How material enters the corpus decides whether retrieval surfaces the right passage, and if it does not, no model rescues the answer.

The index is the product, and it is American#

The Legal Search Index is the new part, so it deserves the most attention.

Per the announcement as quoted by Artificial Lawyer, it searches US case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs, with sources added daily. OpenAI worked with Free Law Project, the nonprofit behind CourtListener, and that collection covers "more than 99.9% of published U.S. precedential case law."

Three things stand out.

It is US-only. Nothing in the coverage I read mentions another jurisdiction or a plan for one. For a US firm that is fine. For everyone else, Astra for Law today is GPT-6 Astra with good legal instructions and no authoritative library for their law.

It leans on public-interest data. CourtListener is the backbone of the case law coverage. That is a sensible choice, and it is also a reminder that the reason a US legal index can be assembled at all is that a nonprofit spent years making American precedent openly available in machine-readable form. The equivalent infrastructure does not exist everywhere.

It does not replace the paid databases. OpenAI says the index "complements the licensed content and specialist products firms rely on from providers such as Thomson Reuters." Thomson Reuters is also among the 26 partner-built plugins launching alongside it, with Intapp, Harvey, Legora, DeepJudge and iManage. The positioning is a platform that the existing legal information business plugs into, not one that displaces it.

Read the benchmark from the other side#

OpenAI tested "Astra for Law's complete setup on 200 U.S. legal research questions from the private validation set of Vals AI's Legal Research Bench." The full setup passed the evaluation's overall correctness check on 54.0% of questions, against 38.7% for GPT-6 Astra using web search alone. That is the source of the "40% more accurate" line in most headlines: it is a relative improvement, 15.3 points on an absolute scale.

On case-law questions it found 24% more reference cases, and it retrieved up to 54% more relevant passages from the correct court opinions.

Two readings of this are worth holding at once.

The first is that it is a genuinely strong result for retrieval. The model did not change between the two runs. The 15-point gain came entirely from the index and the instructions. If you ever needed evidence that in legal AI the retrieval layer is doing the work, OpenAI has now published it.

The second is the complement. On OpenAI's own chosen benchmark, with its own best configuration, 46% of research questions did not pass the overall correctness check. MLQ also points out the comparison is against another OpenAI system, not against the commercial legal research platforms firms already pay for, and Legal IT Insider notes the same absence. The questions come from a private validation set, so nobody outside can rerun them.

None of that makes the result bad. It tells you what kind of tool this is.

Where the verification lands#

The announcement's framing of the output is careful. The goal, as quoted, is "a stronger research foundation for advising on a deal, assessing a dispute, or developing a legal strategy, with reliable authorities the lawyer can examine for herself."

That last clause is the design. Verification sits with the lawyer. The system surfaces authorities; a qualified human checks them. In the coverage I read, there is no description of how citations are attached to claims, and nothing about what the system does when the index holds nothing that supports an answer.

Compare that with the position I set out in no source, no claim. Three rules, in order of how often they get skipped: retrieval gates generation rather than informing it, citations are carried as data rather than requested in a prompt, and refusal is a designed path with its own output.

For a product sold only to supervised legal professionals at large firms, putting the check on the lawyer is defensible. Professional responsibility is the refusal path, moved out of the software and into the person. The risk is what happens when the same configuration reaches the API and gets built into products for people who are not lawyers, or into workflows where nobody has time to examine every authority. At that point the check has to come back into the software, and it has to be structural. A citation the model was asked to produce is a claim about a source. A citation carried through retrieval is a fact about one.

What this means for Saudi and Arabic legal AI#

This is the part I care about most, because it is the part that affects what I build.

Astra for Law does not cover Saudi law. It does not cover any Arabic-language legal system. So for a firm in Riyadh or Dammam, the product as launched is general GPT-6 Astra with legal instructions, which is useful for drafting and reasoning and not a research tool for their jurisdiction.

But the pattern it validates transfers completely, and it points at exactly the work that is hard here.

The corpus is the moat. OpenAI could assemble a US index partly because CourtListener exists. Building the equivalent for Saudi regulation, with each provision in its authoritative Arabic text and its amendment history, is slow and unglamorous work, and it is the work that decides whether answers are right. Whoever does it well owns the result regardless of which model sits on top.

Structure matters more in Arabic. Legal Arabic is precise, formal and structurally specific. Chunking it as though it were prose loses the article and clause boundaries that make a provision meaningful. I wrote about the wider version of this in why Arabic-first software is more than translating an English UI, and legal text is the sharpest case of it.

The model is increasingly interchangeable. If the lesson of Astra for Law is that a frontier model plus a good index beats the same model on the open web by 15 points, then a regional legal platform should be designed so that the model can be swapped as better ones arrive, and the investment should go into the index, the citation plumbing and the refusal path. That is the bet MyAdvisor.ai is built on, and this launch makes me more confident in it, not less.

Data handling is a separate question. For eligible firms the offering includes Zero Data Retention on the API, and ChatGPT Enterprise usage is excluded from human review by default. Those are meaningful controls. They are not the same thing as knowing where data is processed, and for Saudi clients that question still has to be answered on its own terms, against PDPL, rather than inferred from a retention setting.

What I would watch next#

Four things would change how I read this launch.

1. API date and pricing. Neither was published. The API is where Astra for Law stops being a tool for supervised lawyers and becomes a component in other people's products, which is where the verification question gets serious.
2. A second jurisdiction. If the index expands, how it expands matters. Licensed content, public-interest collections like CourtListener, or crawling would each say something different about quality and about who carries the risk of a wrong source.
3. An independent comparison. A result against the commercial research platforms firms already use, on questions outsiders can see, would tell the market far more than a comparison against OpenAI's own web-search baseline.
4. The empty-index behaviour. When the Legal Search Index holds nothing that supports an answer, does the system say so, or does GPT-6 Astra fill the gap from its own parameters? That one design decision is the difference between a research tool and a liability, and it is the thing I would ask OpenAI before anything else.