The product is narrower than the announcement implies
OpenAI’s launch language is ambitious — “frontier intelligence for law,” “a new foundation” — and the underlying reality is more specific. Astra for Law is not a separately trained legal model. It is GPT-6 Astra, configured with a legal search index, legal-writing instructions, and workflow controls aimed at law firms. Independent coverage characterises it as GPT-6 Astra wrapped in legal search and workflow features, not a new base model. That distinction matters before any resourcing or procurement decision.
In ChatGPT it appears as “GPT-6 Astra Law”; the planned API model name is gpt-6-astra-law. Access is currently limited to selected U.S. firms through a Trusted Access programme in ChatGPT and Codex. The API has no published launch date or pricing.
A powerful base model, wrapped in legal tooling, gated behind an invitation programme. That is what you are evaluating.
What the benchmark actually tells you
The headline number is a 54.0% overall correctness score on a 200-question U.S. legal research benchmark, versus 38.7% for GPT-6 Astra using web search alone. The questions came from the private validation set of Vals AI’s Legal Research Bench, run at the highest reasoning effort on both systems. Both configurations use the same underlying model. The 15.3 percentage-point difference measures what the legal search index adds on top of it.
That lift comes from the index, not from any change to GPT-6 Astra itself. The index searches more than 230 million URLs covering U.S. cases, statutes, regulations, court rules, and administrative decisions, with sources added daily.
54% is not a production accuracy rate. Nor is it a comparison against Westlaw, CoCounsel, or any specialist tool your team already uses. GC AI’s analysis is precise on this point: “The result measures the configured research system, not the model alone.” To assess real value, run it against a representative memo or redline from your own practice — not against a private benchmark you cannot inspect.
At 54.0% correctness, 92 of those 200 questions returned an incorrect or incomplete answer at the highest reasoning effort. For first-pass research, that error rate may be workable. For a finalised brief or a privileged opinion, it is not.
Where Astra for Law fits — and where it does not
The strongest case for Astra for Law is volume drafting and early-stage research. John Savva, Partner at Sullivan & Cromwell, noted in his preview that the models “demonstrated impressive research depth and sensitivity to authority” across both litigation and transactional matters. That is a credible signal for research acceleration. It is not a signal that attorney supervision becomes optional.
OpenAI frames the legal search index as complementary to Thomson Reuters and other licensed providers, not a replacement. Thomson Reuters CTO Joel Hron supplied a supportive quote for the launch — on the same day OpenAI started building its own legal research index. The Next Web noted the timing directly. Whether that relationship stays complementary as the index matures is worth watching.
The workflow architecture has practical value independent of the research question. Layer3Labs documents that Astra supports streaming, structured outputs, function calls, and file search, enabling firms to automate routine drafting tasks in custom workflows via API access. Legal IT teams can engineer prompt logging, privilege screens, and integration with document management systems. Those controls, paired with strict access roles and audit trails, add compliance oversight that earlier models made harder to implement.
At launch, 26 partner-built plugins are available, including connections involving iManage, Intapp, DeepJudge, and Thomson Reuters’ HighQ. These are integrations and partner-built tools; OpenAI does not identify those companies as Astra for Law customers. The distinction matters when evaluating vendor claims.
The strongest objection, answered fairly
Your firm likely already pays for Westlaw, CoCounsel, or a comparable specialist product. Those tools are built on curated legal corpora, carry vendor indemnities in some cases, and have years of practitioner feedback incorporated. Astra for Law’s index covers 230 million URLs. Breadth is not the same as depth or curation.
If your research workflow depends on precise citation verification and binding authority ranking, a 54% correctness score on a private benchmark does not displace a specialist tool. A lower capture rate on your own matters would lower the return further — that is a concession worth making explicitly rather than arguing away.
What Astra for Law may do is reduce associate time on first-pass research before a specialist tool does the verification work. That is a narrower value proposition than OpenAI’s framing suggests. It is still a real one.
For data-sensitive matters, GC AI notes that for eligible firms the legal offering includes zero data retention for API usage and an eyes-off human-review posture by default for ChatGPT Enterprise. Confirm those controls are active for your instance before loading privileged documents. Do not assume the default configuration satisfies your jurisdiction’s professional conduct rules.
Three operational decisions to make before deployment
First, decide where in your workflow a 54% first-pass accuracy rate is acceptable. First-pass research, initial drafting, and discovery summarisation are candidates. Final opinions, court submissions, and privileged advice are not — not without a qualified attorney reviewing every output.
Second, if you have API access, build prompt logging and privilege screen infrastructure before broad deployment. Astra’s structured output and function-call support make this tractable. Doing it after deployment creates retroactive compliance exposure.
Third, run your own evaluation before committing budget. GC AI recommends starting with one recurring matter — a vendor agreement, its amendments, and your approved positions — and comparing the research output, the corrections counsel must make, and the total cost. Your firm’s correction rate on that matter is a more reliable deployment threshold than a private benchmark you cannot inspect.
Astra for Law is a meaningful infrastructure step. It is not a replacement for legal judgment, specialist research tools, or the supervision obligations your firm already carries. Used in the right lane, it can accelerate volume work. Used outside it, it creates liability that no benchmark score will cover.
If your firm has or expects Trusted Access: run one representative matter through Astra for Law this week, log every correction a qualified attorney makes, and use that correction rate — not OpenAI’s benchmark — as your deployment threshold.
— Eagentix
Eagentix helps growth-focused enterprises redesign and automate manual business processes. We combine executive strategy, implementation support, and managed services to build dependable operations across Southeast Asia.
Eagentix helps growth-focused enterprises redesign and automate manual business processes. We combine executive strategy, implementation support, and managed services to build dependable operations across Southeast Asia.
Sources
- GPT-6 Astra for Legal Documents | Layer3Labs
- Introducing Astra for Law
- OpenAI’s Astra for Law is a legal configuration of GPT-6 …
- Astra for Law: A Guide for In-House Counsel — GC AI
- Astra for Law: Frontier intelligence built for your practice.
- Astra for Law: Frontier Intelligence Built for Your Practice | OpenAI
- OpenAI launches Astra for Law, and its own legal research index