Treat Astra for Law not as ordinary ChatGPT but as regulated legal infrastructure that OpenAI first offers to selected US firms via Trusted Access. Make explicit choices per matter about which functions you enable, which client data is processed and how every output is logged, reviewed and attributed to a responsible lawyer.
On 17 September 2026 OpenAI announced Astra for Law: GPT-6 Astra tuned to legal work, coupled to a US Legal Search Index and instructions for legal analysis and drafting. Reuters independently confirmed that this is a legally focused version for firms and legal-software vendors. In our assessment this shifts the question for firms from whether they use ChatGPT to how they control this specific, legally tuned tier.
What makes Astra for Law a separate legal ChatGPT tier and not just a model?
According to OpenAI's documentation, Astra for Law would appear as "GPT-6 Astra Law" in ChatGPT's model picker and as gpt-6-astra-law in the API. It is therefore a separately labelled tier, not a hidden setting. LawNext describes it as GPT-6 Astra configured for legal work, with its own legal search index and bespoke instructions, and quotes OpenAI's premise that it supports the expertise and judgement of lawyers rather than replacing them.
The main features according to OpenAI and the reporting:
- A Legal Search Index covering US case law, statutes, regulations and court rules.
- Initial access only for selected US firms via a Trusted Access programme in ChatGPT, with an API version later.
- Zero data retention on the API; ChatGPT Enterprise use is, according to the documentation, excluded from human review by default.
- Collaboration with Latham & Watkins on information permissions, ethical walls, client instructions and firm-level oversight.
That OpenAI itself names governance, privacy and ethics as design premises underlines that this tier is intended for controlled, authorised deployment. Anyone who approaches it the way the rest of managing AI in legal professional practice should be handled misses the point: here the tooling has already been made specific, but responsibility remains with the firm.
Which bar-ethics duties bear directly on the use of Astra for Law?
Layer3 Labs' practical guide translates the use of Astra directly to the ABA Model Rules, in particular Rule 1.6 on confidentiality and Rule 5.3 on responsibility for non-lawyer assistants. The core: treat the AI system as a non-lawyer technological assistant that requires supervision. Layer3 Labs advises using Astra only for intake, drafting and review when the firm deploys non-training accounts (Business, Enterprise, API) and a lawyer reviews every output before it leaves the firm.
In our assessment Astra for Law does not change the ethical floor, but it raises the stakes: because more work can run through AI, the need for demonstrable supervision grows. The question the sources leave open is how that supervision is recorded in practice when a lawyer works via the model picker. Our editorial answer: tie every step to a responsible lawyer and a check point, and capture that in a logging layer as described in demonstrating human oversight of AI decisions.
Which choices must a firm record per matter around Astra for Law?
The source data describes the product, not the set-up per matter; that is an editorial conclusion about where firms now stand. We see the following choices that must be recorded per matter (this is editorial analysis, not a source claim):
- Deciding per matter type whether Astra for Law is enabled, for example for research memos, contract work or regulatory advice.
- Recording which parts of the Legal Search Index are reliable for the relevant area of law.
- Delimiting which client and matter data may be processed, in line with confidentiality duties.
- Registering per use: the model version (GPT-6 Astra Law), the prompts and retrieved sources, the firm instructions applied and the identity and decision of the reviewing lawyer.
Without this registration a firm cannot demonstrate afterwards how Astra influenced a matter. Bounding functions per matter aligns with the broader line of limiting and logging Astra for Law plugins per matter type, and with broader demonstrable control over AI use by lawyers.
Where do the risk boundaries lie: US law, disclosure and multiple tools?
LawNext emphasises that at this stage Astra for Law is aimed at US law and is available on a limited basis via Trusted Access, without published prices. That marks concrete boundaries. From this it follows, as editorial analysis, that firms should therefore:
- avoid deploying Astra on non-US matters, given the scope of the index;
- take account of AI disclosure orders from judges and check per matter whether the use must be reported;
- align Astra with other legal tools such as research platforms, citators and workflow systems, without creating opaque chains of AI decisions.
For legal-tech vendors integrating Astra into products, this means they must make visible which parts of an automated workflow are powered by Astra and under which controls. Astra for Law delivers a powerful but fallible legal tier; it belongs where the index, the controls and human supervision are demonstrable per workflow. That is the essence of treating it as regulated legal infrastructure.
Sources and references
Sources: The article draws on OpenAI's announcement and Help Center documentation about Astra for Law, supplemented by reporting from Reuters and LawNext and the ethics guide from Layer3 Labs.