OpenAI Builds a Legal Configuration of Its Frontier Model, and Puts the Benchmark in Public

OpenAI has introduced Astra for Law, which it describes not as a new model but as a configuration: its GPT-6 Astra model combined with a purpose-built legal search index, instructions for legal analysis and writing, and a set of controls aimed at confidential client work. It is going first to selected law firms through a Trusted Access program in ChatGPT and Codex, with API availability described only as coming soon. The framing matters — OpenAI is positioning this as a foundation that firms and legal software companies build on, not a product that replaces what they already run.

What is in the box

Three pieces make up the configuration. The first is the model, which OpenAI calls its latest and most powerful. The second is a new legal search index covering US case law, statutes, regulations, court rules and administrative decisions across what the company says is a corpus of more than 230 million URLs, with sources added daily. OpenAI credits work with the Free Law Project, the nonprofit behind CourtListener, for bringing in a case-law collection it says covers more than 99.9% of published US precedential case law. The third piece is a set of custom instructions meant to govern how the model reasons and writes — distinguishing a holding from surrounding commentary, addressing authority that cuts against an argument, explaining how a contract exception shifts risk.

It will appear in the ChatGPT model picker as "GPT-6 Astra Law" and, when the API opens, under a matching model identifier. OpenAI says it plans to bring these legal capabilities forward to its newer frontier models as they arrive. Pricing was not disclosed.

The benchmark, and how to read it

OpenAI published numbers rather than adjectives, which is worth acknowledging. It tested the full Astra for Law setup against 200 US legal research questions drawn from the private validation set of Vals AI's Legal Research Bench, a benchmark that scores whether a system finds the right sources and passages and whether its answers meet the evaluation criteria.

At the highest reasoning effort for both systems, OpenAI reports Astra for Law passed the overall correctness check on 54.0% of questions, against 38.7% for GPT-6 Astra using web search alone — a 40% relative improvement. On case-law questions it says the configuration found 24% more reference cases, and on an audited set of target passages retrieved up to 54% more relevant passages from the correct opinions.

The gap is real and the direction is clear: a licensed, structured legal corpus beats general web search for this work. But 54.0% is the number a partner should sit with. On a benchmark of research questions, the configured system still fails the overall correctness check on nearly half. That is a research assistant whose output requires verification on every use, not a system anyone should treat as settled authority. OpenAI's own framing gestures at this, describing the value as producing authorities the lawyer can examine herself. The comparison is also OpenAI's own, run on a private validation set, against its own baseline — the kind of result that gains credibility only when third parties reproduce it.

OpenAI also published side-by-side outputs against a competing frontier model on a litigation and a transactional prompt, noting that the rival returned a holding that had been reversed on appeal in one example and found no matching case in the other. Two prompts chosen by the vendor are an illustration, not a study.

Confidentiality is the actual gate

Law firms do not adopt tools on benchmark scores. They adopt them when the confidentiality story survives a general counsel and an ethics committee, and OpenAI has clearly built the offering around that.

The Trusted Access Program is restricted to eligible firms, for lawyers and people working under their supervision, on professional legal work. OpenAI says it includes Zero Data Retention on the API, and that ChatGPT Enterprise usage under the program is excluded from human review by default. The company also says it is working with Latham & Watkins to design for information permissions, ethical walls, client instructions and firm oversight — the machinery that keeps one matter's data from reaching a team conflicted out of it.

That last item is the hard part and the least specified. Ethical walls are not a settings toggle; they are enforced across document systems, time entries, email and now a model's retrieval scope. OpenAI describes the work as in progress with a design partner, not as a shipped capability with published guarantees. Eligibility criteria for the program were not detailed, and neither was what happens to firms outside it.

An ecosystem play, deliberately

The more interesting strategic choice is how much OpenAI is not building. API customers including Harvey and Legora will be able to build on Astra for Law inside their own products — meaning two prominent legal AI companies become distribution rather than competition, at least for now. OpenAI is launching 26 partner-built plugins connecting ChatGPT to tools firms already run, naming Relativity, Clio, iManage, Intapp and DeepJudge, with Thomson Reuters bringing HighQ matter context into ChatGPT and previewing a forthcoming CoCounsel Legal connector. Nine community plugins from legal engineering groups add 47 custom skills practitioners can adapt. ChatGPT for Word became generally available the same day.

OpenAI also describes forward-deployed engineers adapting ChatGPT Enterprise for individual firms: Sullivan & Cromwell built an agreement analyzer that brings negotiating playbooks into deal review and turns findings into proposed redlines; Ropes & Gray built a diligence system around how its lawyers work a data room; Cooley built a product for IPO preparation; Skadden is designing regulatory-risk tooling; Wachtell, Lipton, Rosen & Katz is collaborating on litigation and corporate work.

Read together, this is a platform strategy aimed at the layer where legal software vendors have historically captured margin. OpenAI supplies frontier intelligence, retrieval and controls; firms and vendors supply expertise and workflow. It is a generous posture while OpenAI needs the ecosystem's credibility and data access. Whether it stays generous once firms have built their workflows on the platform is the question every partner in that list is presumably modeling.

The quotes OpenAI published are correspondingly measured. A Sullivan & Cromwell partner calls the previews a significant step and describes research depth and sensitivity to authority. Harvey's applied research lead cites grounding and citation precision. Nobody claims the work is finished.

What to watch

• Whether any independent party reproduces the Legal Research Bench result, and whether OpenAI publishes results at lower reasoning effort settings where cost and latency are practical.
• How the remaining error rate is handled in practice — whether firms build mandatory citation-verification steps, and whether malpractice insurers take a view.
• The specifics of ethical-wall enforcement once the Latham & Watkins collaboration produces something shippable, rather than a design partnership.
• Whether Harvey, Legora and Thomson Reuters remain ecosystem partners as OpenAI's own firm-facing engineering work expands.
• Access and price beyond selected firms. Nothing announced addresses small firms, solo practitioners or in-house teams, where the volume of routine legal work sits.