OpenAI's legal index lifted GPT-6 Astra from 38.7% to 54% on 200 bench questions

An associate at one of the selected firms opens the model picker in ChatGPT and finds a new entry, GPT-6 Astra Law ^[9]. The weights are the same GPT-6 Astra that began rolling out on Sept. 3 ^[20]. The benchmark OpenAI chose rewards finding the right authority and the right passage inside it ^[2], and on that test the indexed configuration beat the plain one by 15.3 points ^[22].

That gap is the product. Nothing was retrained; the model got a search index and a set of instructions ^[1]. OpenAI says the configuration surfaced 24% more reference cases on case-law questions, and on a separate audited set retrieved up to 54% more of the target passages from the correct opinions ^[8]. It still got 92 of the 200 questions wrong ^[23]. Wrong answers fell from 61.3% to 46%, about a quarter of the failures ^[24].

The baseline figure does not divide. 38.7% of 200 questions is 77.4 questions; 77 would be 38.5% and 78 would be 39% ^[25]. OpenAI does not say what the denominator was.

The buying decision happens a level up from the associate. Harvey and Legora, two of the companies that sell legal AI into firms, are named as API customers for gpt-6-astra-law ^[11]. So the retrieval advantage a vendor could present as its own engineering becomes a line item those vendors rent, at a rate OpenAI has not disclosed ^[10]. OpenAI called the launch the start of a long-term investment in law ^[21].

Thomson Reuters is bringing HighQ matter context into ChatGPT and previewing a connector for its CoCounsel Legal product ^[4]. "As AI becomes more open and interoperable, the value is not in connectivity alone," said Joel Hron, the company's chief technology officer. "Legal professionals need more than access to information. They need trusted intelligence, relevant enterprise and matter context, purpose built legal capabilities, and the governance required for high stakes work." ^[5]

Some of that governance is still being drawn. Latham & Watkins is working with OpenAI on governance design for information permissions, ethical walls and client instructions ^[13]. Eligible firms get zero data retention on the API, and ChatGPT Enterprise usage is excluded from human review by default ^[12]. A firm holding both sides of a matter will want the wall specified before the pilot starts.

OpenAI engineers embedded at individual firms have been building on ChatGPT Enterprise. Sullivan & Cromwell has an agreement analyzer that pulls its negotiating playbooks and selected precedents into the review of a new deal and turns what it finds into proposed redlines ^[17]. At Ropes & Gray the work went into deal diligence, and Cooley's tool, GO Public, handles initial public offering preparation, drafting the filing included ^[18]. John Savva, a Sullivan & Cromwell partner who saw early builds, said the models showed "impressive research depth and sensitivity to authority" ^[19].

The index is one thing on sale in this launch, and OpenAI will rent it to anyone once the API has a price. Matter context is another: a firm's own documents, permissions and ethical walls. The plugin set is reaching for that ^[14][15]. The third is the workflow, where a redline or a draft filing comes out. Every line on a vendor invoice belongs in one of the three. Anything in the first column is resold gpt-6-astra-law, at a markup only OpenAI can see today.