Product and intended use
Define the configuration, people, information and tasks being assessed.
Product and plan covered
Public claim only
Astra for Law is described as GPT-6 Astra combined with a Legal Search Index, custom instructions for legal analysis and writing, and settings for professional legal work. The help article, checked 2026-09-27, states it 'will initially be offered to selected law firms in the United States through Trusted Access, Codex, and will be coming soon to the API', and that access is 'for eligible lawyers and people working under their supervision'. It appears as 'GPT-6 Astra Law' in the ChatGPT model picker and as gpt-6-astra-law in the API.
Ask: Is Astra for Law available to an Australian-domiciled firm today, on which order form and under which OpenAI entity, and what are the written Trusted Access eligibility criteria?
Deployment model
Public claim only
Delivery is vendor-hosted only: through ChatGPT and Codex under Trusted Access, with API availability stated as coming soon. No private cloud, on-premises or self-hosted option is described on either the announcement or the law solutions page, both read 2026-09-27.
Ask: Is any dedicated-tenancy or in-region deployment available for Astra for Law, or is ChatGPT Enterprise multi-tenant SaaS the only option?
What the product does
Public claim only
Three components are described: the Legal Search Index as a tool the model can call over US primary law; custom instructions that guide applying research to a client's facts, developing arguments or deal terms, and identifying weaknesses and uncertainty; and settings for professional legal work. The announcement also describes 26 partner plugins, and separately describes three applications that named US firms built on ChatGPT Enterprise with OpenAI engineers, which are not presented as features of Astra for Law.
Ask: Which of the published components are in the Trusted Access offering itself, and which require separate engineering work or a separate plugin subscription?
Jurisdictional coverage of the legal corpus
Public claim only
The Legal Search Index is described as covering 'U.S. case law, statutes, regulations, court rules, and administrative decisions' across a corpus of more than 230 million URLs with sources added daily, incorporating the Free Law Project's CourtListener collection. No Australian or other non-US legal corpus is named on the announcement, the law solutions page or the help article, all read 2026-09-27. The published benchmark questions are also all US law.
Ask: Does the Legal Search Index contain any Australian legislation or case law, and if not, what does the model rely on when asked an Australian law question inside Astra for Law?
Vendor content directing how AI describes the product
Not confirmed
Checked deliberately on 2026-09-27. No page or block directing how AI models, assistants or crawlers should describe the product was found on the announcement, the law solutions page or the Astra for Law help article, and the site map taken in the previous phase (38 links from openai.com) surfaced no /llms.txt, /ai-info or equivalent path. The announcement does carry vendor framing that is not adopted here, including 'frontier intelligence for law' and 'legal-grade trust and controls'.
Ask: Does OpenAI publish any machine-directed guidance (for example llms.txt or an AI-information page) about how Astra for Law should be described, and where?
Data, controls and requirements
Compare the actual configuration with firm, client and jurisdiction requirements, supported by evidence.
Client authority and professional responsibility
Public claim only
Services Agreement clause 4.3 puts the customer 'solely responsible for all use of the Outputs and for evaluating the accuracy and appropriateness of Output for Customer's use case', and clause 3.2 requires the customer to obtain and maintain any end-user consents necessary. Access is described as limited to eligible lawyers and people working under their supervision. There is no clause about client confidences, ethical walls or client instructions in the agreement read; the announcement describes design work with a US firm on 'information permissions, ethical walls, client instructions, and firm oversight' as work in progress.
Ask: What client-consent and ethical-wall controls exist in the product today, as opposed to the design work described in the announcement, and are any of them contractual?
Foundation-model provider
Public claim only
The model is the vendor's own: GPT-6 Astra, presented as 'GPT-6 Astra Law' in the ChatGPT model picker and gpt-6-astra-law in the API. No third-party model provider is named, and no customer model choice is described for this configuration; the announcement states that as its frontier models advance it will 'bring these legal capabilities to our latest models'.
Ask: What notice will we receive before the underlying model behind Astra for Law changes, and can we pin a model snapshot for a matter that is under way?
Subprocessors
Not confirmed
No subprocessor list was opened in this pass and the enterprise privacy page read on 2026-09-27 does not link one. Services Agreement clause 9.2 extends the customer's limited permission to 'subcontractors or sub-processors', and clause 5.3 incorporates the DPA, which the definitions publish at https://openai.com/policies/data-processing-addendum; neither document was read here. Credit budget was spent on the product help article, the enterprise privacy page, the Services Agreement and the residency article instead.
Ask: Provide the current dated subprocessor register applying to Astra for Law, including any provider used for the Legal Search Index, with entity, country, purpose and our notice and objection rights.
Training on customer content
Public claim only
Services Agreement clause 4.2: 'OpenAI will not use Customer Content to develop or improve the Services, unless Customer explicitly agrees to such use.' The enterprise privacy page (updated 8 January 2026) states models are not trained on business data by default, that data from ChatGPT Business, Enterprise, Edu, Teachers, Healthcare and the API Platform after 1 March 2023 is not used for training unless the customer has explicitly opted in, and that data accessed from apps is not used for training by default.
Ask: Confirm in writing that clause 4.2 covers prompts, documents, outputs, embeddings and Legal Search Index queries for our workspace, and that no opt-in has been enabled for any of our users.
Feedback and quality monitoring
Public claim only
Services Agreement clause 9.3: if the customer provides Feedback, OpenAI may 'use and exploit Feedback without restriction or compensation'. The enterprise privacy page states that if a customer has explicitly opted in to share data through feedback mechanisms to improve the services, the shared data may be used to train models, and that business data may be run through automated content classifiers whose classifications are metadata rather than the business data itself.
Ask: Can in-product feedback and model-feedback sharing be disabled for the whole workspace, and does any feedback path carry matter text out of our retention settings?
Retention of prompts, outputs and logs
Public claim only
Retention depends on which surface is used. For ChatGPT Enterprise the enterprise privacy page says workspace admins control how long data is retained and deleted conversations are removed within thirty days unless legally required. For the API it says inputs and outputs may be retained up to thirty days for abuse monitoring, except for endpoints and features listed in the platform documentation, with zero data retention available for eligible endpoints and qualifying use cases; the vendor states the legal offering includes Zero Data Retention on the API. No retention period is published for Legal Search Index queries specifically.
Ask: State the default and configurable retention for Astra for Law prompts, outputs, uploaded files and Legal Search Index queries, and whether Zero Data Retention applies to the ChatGPT surface as well as the API.
Deletion and certified erasure
Public claim only
Services Agreement clause 11.3: on termination 'OpenAI will delete all Customer Content from its systems within thirty days' unless legally required to retain it or the customer has agreed otherwise in writing, with an express exception allowing OpenAI to retain Abusive Customer Content and, where access is suspended or terminated, to share information related to it. The enterprise privacy page states deleted conversations are removed within thirty days. No certified-erasure statement was found.
Ask: Will you certify deletion in writing within the thirty days, and does the certificate cover backups, abuse-monitoring logs and any subprocessor copies?
Vendor and contractor access to content
Public claim only
The enterprise privacy page states that for ChatGPT Enterprise, Edu and Healthcare, 'Authorized OpenAI employees will only ever access your conversations for the purposes of resolving incidents, recovering end user conversations with your explicit permission, or where required by applicable law', while for ChatGPT Business, Teachers and the API it also names 'specialized third-party contractors' bound by confidentiality who review for abuse and misuse. It also states business data may be run through automated classifiers. The law solutions page states 'ChatGPT Enterprise usage comes with an “eyes-off” human review posture by default' for the legal offering.
Ask: Is the eyes-off human review posture a contractual commitment for our workspace or a default setting, and can any OpenAI employee or contractor read Astra for Law conversations without a ticketed request we can audit?
Access control, SSO and audit
Public claim only
The enterprise privacy page lists enterprise-level authentication through SAML SSO, fine-grained control over access and available features, admin control over which internal sources are connected, and an audit log of conversations and GPTs available to workspace admins through the Enterprise Compliance API. Services Agreement clause 3.1 prohibits shared credentials and requires prompt notice of unauthorised access. No role model specific to Astra for Law is described.
Ask: Can access to Astra for Law be restricted to nominated practice groups, is multi-factor authentication enforceable, and does the Compliance API export Legal Search Index queries?
Encryption in transit and at rest
Public claim only
The enterprise privacy page states data encryption at rest with AES-256 and in transit between customers and OpenAI and between OpenAI and its service providers with TLS 1.2 or above, with strict access controls. Services Agreement clause 5.1 binds OpenAI to published Security Measures and gives the customer a right to terminate within five business days if an update materially diminishes them; the Security Measures document itself was not opened in this pass. Customer-managed keys are not mentioned.
Ask: Are customer-managed or customer-held encryption keys available, and what do the Security Measures say about key management for the workspace holding our matter content?
Independent security assurance
Public claim only
The enterprise privacy page states ChatGPT Enterprise, Edu and Healthcare have each completed a SOC 2 Type 2 audit and that the API Platform is SOC 2 Type 2 certified, and points to the Trust Portal at trust.openai.com, which was not opened in this pass. Services Agreement clause 5.2 gives the customer a right, on written request and no more than once a year, to a copy of the most recent Audit Reports, which are deemed OpenAI Confidential Information.
Ask: Provide the current SOC 2 Type 2 report and confirm whether its scope covers Astra for Law and the Legal Search Index, plus any ISO 27001 or ISO/IEC 42001 certificate and the latest penetration-test summary.
Contracting entity, governing law, forum and liability
Public claim only
The Services Agreement, effective 1 January 2026 and read 2026-09-27, defines the 'OpenAI Contracting Party' as OpenAI OpCo, LLC for customers located outside the EEA or Switzerland; OpenAI Ireland Ltd applies to EEA and Swiss customers only. For all customers outside the EEA, Switzerland and the UK the Governing Laws are the laws of California and the Venue is the federal or state courts of San Francisco County, and clause 15 requires final and binding individual arbitration through NAM with no class actions. Liability is capped at amounts paid in the preceding twelve months, with carve-outs for gross negligence or wilful misconduct, indemnities and payment obligations. The DPA is incorporated by clause 5.3. No Australian entity, ACN or ABN appears.
Ask: Will you contract through an Australian entity, accept Australian law and an Australian forum, and disapply the arbitration and class-action waiver for an Australian law firm?
Cross-border transfers, Australian residency and inference location
Public claim only
The vendor's residency article, read 2026-09-27, lists ChatGPT data residency at rest as available in Australia, Canada, Europe (EEA and Switzerland), India, Japan, Singapore, South Korea, the United Arab Emirates, the United Kingdom and the United States, for eligible new ChatGPT Enterprise and Edu customers and eligible API customers. Inference residency is available only for Europe, the United States and the United Arab Emirates, so GPU execution on Australian-stored content is not kept in Australia. The article also states that features which are not generally available are not eligible for data or inference residency, that external integrations including Apps, MCP and Web Search are handled under the other provider's terms and may sit outside the region, and that workspace metadata, billing information and user logins may be stored outside it.
Ask: Astra for Law is offered through Trusted Access rather than as a generally available feature: is it eligible for Australian data residency at all, and in which country does its inference and Legal Search Index retrieval actually run?
Incident notification and history
Not confirmed
No notification period was found on the pages read. The Services Agreement addresses suspension for a 'Security Emergency' and requires the customer to notify OpenAI of unauthorised account access, but contains no breach-notification clause; the enterprise privacy page describes a security team with 24/7/365 on-call paging and a bug bounty programme. The notification terms would sit in the DPA, which is incorporated by clause 5.3 and was not opened in this pass.
Ask: State the contractual notification period from discovery for our workspace, whether it covers subprocessor incidents, and how you would support a notifiable data breach assessment under Australian law.
Change notice for terms, models and features
Public claim only
Services Agreement clause 16.13: OpenAI may update the agreement or its policies with reasonable notice including by posting, and will give at least thirty days notice where it determines in its sole judgement that an update materially impacts the customer's rights, except where an update is needed to comply with law; continued use constitutes acceptance. Clause 2.3 allows termination on thirty days notice if an update materially reduces functionality, exercised within five business days of notice, and that right does not apply to beta or evaluation features. Price changes on the pricing page take effect fourteen days after posting. The announcement states legal capabilities will be brought to later frontier models.
Ask: What notice will we get before the model, the Legal Search Index contents or the legal instructions behind Astra for Law change, and is Astra for Law a beta or evaluation feature for the purposes of clauses 2.3 and 12.3?
Export and exit
Public claim only
Services Agreement clause 11.2 allows either party to terminate for uncured material breach after thirty days or on insolvency; clause 11.3 ends the customer's rights immediately on termination and requires deletion of Customer Content within thirty days, and makes unpaid minimum commitment amounts immediately due except where the customer terminates for cause. Prepaid fees are non-refundable except as required by law. The enterprise privacy page describes the Enterprise Compliance API for conversation and GPT audit access; no export format or bulk-export commitment for Astra for Law work product is published.
Ask: Before termination, in what format can we export conversations, generated memoranda and citation trails, and for how long will that export remain available?
Connectors, plugins and third-party services
Public claim only
Services Agreement clause 3.4 makes Third-Party Services optional and subject to that provider's own Third-Party Service Terms as well as the agreement. The enterprise privacy page states workspace admins control which apps are enabled, ChatGPT respects existing permissions, each end user must authenticate with the connected application, and data accessed from apps is not used for training by default. The residency article states information sent to an external app, MCP server or web browsing provider 'is handled under that provider's own storage, processing, privacy, and data-residency terms'.
Ask: Which plugins would be enabled for our workspace, what matter content does each receive, and will you confirm in writing that none are on unless we enable them?
Open-web retrieval
Public claim only
Web search is a separate, switchable ChatGPT feature rather than part of the Legal Search Index: the announcement's benchmark compares Astra for Law against 'GPT-6 Astra using web search alone', and the residency article lists 'Web Search, if enabled' among external integrations whose data may be stored and processed outside the selected region under the other provider's terms. Whether web search remains available inside an Astra for Law session, and whether it can be disabled independently, is not published.
Ask: In an Astra for Law session, can the model reach the open web as well as the Legal Search Index, which browsing provider is used, and can web search be switched off for the workspace while Astra for Law stays on?
Agent actions and confirmation
Not confirmed
Astra for Law calls the Legal Search Index as a tool and is offered in Codex, and the announcement describes plugin actions that write to other systems, for example that 'with iManage, a lawyer can draft a negotiation brief in ChatGPT and save it to the matter file' and that Intapp 'can surface activities that may need a time entry for review'. No document read describes a confirmation step, an action allow-list or an action audit trail for those writes.
Ask: Which plugin actions can write to our document or practice management system without a separate confirmation, and where is the audit record of each action kept?
Human review before use
Public claim only
The help article instructs: 'Review the answers and cited sources before relying on them.' Services Agreement clause 4.3 makes the customer solely responsible for all use of Outputs and for evaluating their accuracy and appropriateness, and clause 12.2 disclaims any warranty that Customer Content will be accurate. Neither document describes a product control that forces a named reviewer before an output leaves the firm.
Ask: What product controls, as opposed to instructions, can we configure to require and record a lawyer's review of an Astra for Law output before it is sent to a client?
