OpenAI President Greg Brockman: Plugin for Law Firms Launched to Significantly Accelerate Law Firm Operations
OpenAI President Greg Brockman shared a company announcement stating that the plugin for law firms, Astra for Law, has been launched to significantly accelerate law firm operations while making data privacy a core feature. ChatGPT is equipped with 26 partner plugins and 47 community plugins.
The product is a legal configuration of GPT-6 Astra, featuring legal search indexing and instructions for analysis and writing, serving as a foundation for law firms and legal tech companies, rather than just a chat window. The index covers U.S. case law, regulations, court rules, and administrative decisions, with over 230 million URLs, and collaborates with the Free Law Project's CourtListener to access over 99.9% of published U.S. precedents. It is displayed as GPT-6 Astra Law in the model selector, with the API named gpt-6-astra-law.
The initial access path is through trusted access: selected law firms can use it in ChatGPT and Codex, with the API following suit. API clients like Harvey and Legora can embed the same intelligence into their own products. OpenAI specifically mentioned collaborations with Cooley, Ropes & Gray, and Sullivan & Cromwell to create workflows using ChatGPT Enterprise, and is exploring judgment support in litigation and corporate business with Wachtell, Lipton, Rosen & Katz. Target clients include AMLaw 200.
The plugin layer integrates existing legal software into the conversation. The 26 vendor plugins cover Relativity, Clio, iManage, Intapp, DeepJudge, Thomson Reuters, Harvey, Legora, etc.: iManage can complete and save negotiation memos back to matters; Intapp suggests activities that may need time tracking; DeepJudge pulls up old transactions for comparison; Thomson Reuters sends HighQ case details into ChatGPT and previews the CoCounsel legal connector. Additionally, there are 9 community plugins developed by legal professionals and engineers, offering 47 customizable skills.
In evaluations, OpenAI used a private validation set of 200 U.S. questions from the Vals AI legal research benchmark: under maximum reasoning strength, Astra for Law achieved an overall accuracy of 54.0%, while GPT-6 Astra using only web search scored 38.7%, representing an approximate 40% improvement; for case questions, it retrieved 24% more reference cases and extracted up to 54% more relevant paragraphs from correct judgment opinions under the same reasoning strength. The company stated that the answer length is about twice that of regular Astra, due to lawyer feedback requesting longer texts. Privacy measures include legal-grade controls, zero data retention, and ChatGPT Enterprise usage not being subject to human review.
The market mechanism involves model vendors competing for legal workflow entry points. Buyers are large firms and legal tech vendors that must maintain confidentiality obligations, trading trusted access for cutting-edge models; sellers are OpenAI, packaging the most expensive models, search indexes, and plugin ecosystems. Beneficiaries are platforms that can integrate search, document systems, and time tracking software into the same conversation, as well as legal AIs like Harvey and Legora that can resell models; those under pressure are general models relying solely on web search and outdated legal search databases that refuse to open interfaces. Funding flows are shifting towards reasoning and plugin distribution based on case billing, rather than purchasing isolated legal database accounts.
Source: Public Information
ABAB AI Insight
OpenAI's transition from Codex writing code to industry models handling professional work mirrors Microsoft's transformation of Office into a platform: first making general models sufficiently functional, then incorporating high-fee vertical scenarios into its distribution layer. The legal industry, characterized by hourly billing, high costs of errors, and absolute confidentiality of client documents, is precisely caught at the intersection of "models must be strong, data cannot leak, and outputs must be citable." Brockman emphasizes privacy as a core feature because large firms will not hand over client files to consumer-grade chatbots capable of human review. Collaborating with Sullivan & Cromwell, Ropes & Gray, Cooley, and Wachtell is aimed at aligning the model with workflows from top firms, rather than waiting for law firms to write prompts themselves.
The capital path involves dual-layer rent-seeking. The base layer sells GPT-6 Astra Law to Harvey and Legora, allowing them to continue serving lawyers; the upper layer integrates 26 vendor plugins and 47 community skills into ChatGPT, enabling firms that have already purchased Relativity, Clio, and iManage to embed the model into matters without changing systems. OpenAI simultaneously builds a legal index of 230 million URLs and connects to CourtListener, reducing reliance on Thomson Reuters and LexisNexis full-text databases. Money is shifting from "buying databases by seat" to "billing by reasoning calls plus plugin distribution"; while Anthropic focuses on direct connections with large firms and security narratives, OpenAI counters with a plugin ecosystem and trusted access for AMLaw 200.
Key comparisons should be noted. Thomson Reuters' Westlaw and CoCounsel are old retrieval aristocrats; Harvey packages large models into a lawyer's workstation; Legora has already run 41 materials for tie-out using Astra in transaction document verification. OpenAI now supplies models to Harvey while also placing Harvey plugins in ChatGPT, effectively acting as both an arms dealer and a shelf. The industry phase is shifting from expanding demonstration clients to controlling workflows: whoever masters case indexing, matter write-backs, and zero retention will capture the layer lawyers open daily.
Structural judgments indicate a reconstruction of the industry chain. Legal knowledge production has transitioned from "database companies selling retrieval + law firms billing hours" to integrating search indexes, document systems, time tracking software, and reasoning models into the same conversation. The mechanism requires that hallucinations must be constrained by citable materials, and confidentiality must be safeguarded by zero retention and trusted access, compelling general model vendors to develop their own legal corpus and plugin standards. Pricing power is partially shifting from hourly partners to platforms that can reliably cite precedents and write back to case file systems.
ABAB News · Cognitive Law
The threshold for professional work is citation and confidentiality, not chatting.
Whoever controls the interface for writing back to case files controls the workflow.
For general models to penetrate high-fee industries, they must first build their own corpus and permissions.