OpenAI's new legal AI is already redlining contracts at Sullivan & Cromwell. Here's what that means for your next subcontract
Astra for Law pairs GPT-6 with a 230-million-URL case law index and a 40% jump in legal-research accuracy, and a top firm is already using it to redline agreements against its own playbook. That's the same clause-by-clause review a trade sub usually skips to save money.
OpenAI launched Astra for Law on September 17 — a version of GPT-6 Astra wrapped in a 230-million-URL legal search index — and Sullivan & Cromwell is already using it to redline contracts against the firm's own negotiating playbook, flag risk clauses, and draft client guidance automatically. That's the exact task a trade sub usually skips: full clause-by-clause review of a subcontract before signing it.
What did OpenAI actually ship?
Astra for Law combines three things: the base GPT-6 Astra model, a Legal Search Index that reaches U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs (built on the Free Law Project's CourtListener database, which covers 99.9% of published U.S. precedential case law), and instructions tuned for legal reasoning and drafting. On Vals AI's Legal Research Bench — a 200-question validation set — Astra for Law passed the overall correctness check on 54% of questions, versus 38.7% for GPT-6 Astra using ordinary web search. That's a 40% relative jump, and it found 24% more reference cases and pulled more relevant passages from the right opinions.
It's rolling out through a Trusted Access program inside ChatGPT and Codex, currently limited to select firms — Latham & Watkins, Ropes & Gray, Cooley, and Sullivan & Cromwell among them — with an API version and 26 partner integrations (Relativity, Clio, iManage, DeepJudge) coming later.
Why does a BigLaw research tool matter to a sub signing a subcontract?
Because the one piece that stands out isn't the benchmark score — it's what Sullivan & Cromwell built with it. The firm's agreement-analysis application feeds contracts through Astra for Law against its own playbook and precedent library, then proposes redlines and risk flags automatically. That's the same review a trade sub's subcontract usually needs and usually doesn't get, because a full attorney redline costs real money and most subs don't have the negotiating weight — or the time before mobilization — to push back on every clause. The clauses that get glossed over under deadline pressure are consistently the ones that decide who eats a loss later:
| Clause | What a full redline catches |
|---|---|
| Indemnification | Whether you're covering the GC's own negligence, not just your own work |
| Pay-if-paid vs. pay-when-paid | Whether your payment is contingent on the owner paying the GC at all |
| Liquidated damages | Whether the LD rate is tied to your scope or the whole project's schedule |
| Termination for convenience | Whether you get paid for work in progress and demobilization costs |
| Assignment of subcontract | Whether the GC can hand your contract to another party without your consent |
What should a sub actually do with this?
Nothing changes today — Astra for Law isn't available to a construction-focused firm yet, and even inside the firms that have it, a lawyer still has to verify the output. But this is the pattern to watch, not the product to buy: BigLaw contract-review tooling has historically taken one to two years to show up in a cheaper form at mid-market and boutique firms, the same path e-discovery and contract-analytics software took a decade ago. The practical move now is to ask whatever counsel reviews your subcontracts — in-house, a GC's risk team, or outside counsel — whether they're tracking tools like this, and to keep pushing for full clause review on indemnification, pay-if-paid, and liquidated damages language regardless of how the review gets done. The bottleneck OpenAI just showed a path around is cost and turnaround time on legal research, not judgment — and judgment is still what decides whether a redline actually protects you.
Friday one chart. Every week, one piece of data that should change a decision on your project. Subscribe at constructionaibrief.com.
- What is OpenAI's Astra for Law?
- It's a legal-research configuration of GPT-6 Astra that OpenAI launched on September 17, 2026. It pairs the model with a Legal Search Index covering more than 230 million URLs of U.S. case law, statutes, regulations, court rules, and administrative decisions, plus instructions tuned for legal analysis and drafting.
- Does this replace a construction lawyer for reviewing a subcontract?
- No. It's currently limited to a Trusted Access rollout inside a handful of large law firms, still requires a licensed attorney to check its output, and passed only 54% of a 200-question legal research benchmark — well short of reliable on its own. Jurisdiction-specific issues like mechanic's lien deadlines and prompt-payment act interpretation still need a lawyer licensed in that state.
- Who is already using Astra for Law?
- OpenAI named Latham & Watkins, Ropes & Gray, Cooley, and Sullivan & Cromwell as early adopters. Sullivan & Cromwell built an agreement-analysis application on top of it that checks contracts against the firm's own negotiating playbook and precedent, flags risk, and proposes redlines.
- Will this make subcontract review cheaper for a trade sub?
- Not immediately — access is restricted to large firms today. But the same clause-by-clause redline capability Sullivan & Cromwell built for its own playbook is the kind of tool that construction-focused firms and legal-tech vendors typically build a cheaper version of within a year or two, the way e-discovery and contract-analytics tools moved from BigLaw to mid-market practices.
- Does the case-law index help with mechanic's lien or delay-claim research?
- The index draws on the Free Law Project's CourtListener database, which covers 99.9% of published U.S. precedential case law, so it has the raw material for state-specific lien and delay-claim precedent. OpenAI hasn't shipped a construction-specific module, so today it's general-purpose legal research, not a lien-law tool.