OpenAI's new legal configuration of GPT-6 Astra costs more per token than the base model, and Cooley's own case study shows what the tool is actually meant to replace.
An IPO can turn a law firm's capital markets team into a document-sorting operation for weeks before anyone reaches the parts of the deal that actually need a lawyer's judgment.
"When there's an IPO, there are thousands of things that need to be done constantly," says David Wang, Chief Innovation Officer at Cooley, in OpenAI's case study on the firm's GO Public tool. Cooley built GO Public to handle that first pass, so its lawyers can spend more time on the judgment calls the software should not be making.
On September 17, OpenAI folded the model behind that tool into a wider offering it calls Astra for Law: GPT-6 Astra, its newest model, wrapped in a legal search index and a set of instructions built for legal research, analysis, and drafting.
What changed
The new legal search index reaches U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, refreshed daily. A partnership with the nonprofit Free Law Project brings in its CourtListener database, which OpenAI says covers more than 99.9% of published U.S. precedential case law.
OpenAI tested the configuration against 200 questions from Vals AI's private Legal Research Bench validation set. Astra for Law passed the benchmark's overall correctness check on 54.0% of questions, versus 38.7% for GPT-6 Astra using ordinary web search, a 40% relative improvement. It also found 24% more reference cases on case-law questions and retrieved up to 54% more relevant passages from the correct court opinions, figures independently reported by SiliconANGLE.
Access is narrow for now. Astra for Law is rolling out through a Trusted Access Program to selected U.S. law firms inside ChatGPT and Codex, appearing in the model picker as "GPT-6 Astra Law." An API version, gpt-6-astra-law, is coming later with no firm date attached, according to Artificial Lawyer. Sullivan & Cromwell, Ropes & Gray, Cooley, and Latham & Watkins are already building on early versions, alongside API partners Harvey and Legora. OpenAI also shipped 26 partner plugins connecting tools like Relativity, Clio, and Thomson Reuters' HighQ, plus 47 community-built skills, and made ChatGPT for Word generally available the same day.
Why it matters
For a firm's finance team, the number worth knowing sits in OpenAI's own enterprise rate card: GPT-6 Astra Law is priced at $12.50 per million input tokens and $62.50 per million output tokens on token-based Enterprise agreements, above the $10.00 and $50.00 rates for the general-purpose GPT-6 Astra it's built on. Intelligence tuned for legal work costs more to run than the model underneath it, and that premium sits on top of whatever a firm already pays for its Enterprise or Work plan.
That pricing lands in an industry that is already comfortable raising its own rates. The Thomson Reuters Institute's Law Firm Rates Report 2026 found that law firm billing rates rose 7.4% in 2025, more than double the 2.8% inflation rate, and warned that clients are growing more cost-conscious about it. A tool that gets a first draft of an IPO filing or a diligence memo to a stronger starting point sooner is one way a firm might defend that rate line. It's also one more line item a cost-conscious client may ask about at the next renewal.
The honest tension
None of this promises a faster verdict on whether a claim actually holds up. Even at its best, Astra for Law's research still failed the correctness check on close to half the benchmark's questions. Cooley is careful about what the tool is actually meant to replace. "The point isn't simply to do the same work faster," says Dave Peinsipp, the firm's capital markets co-chair, in the same case study. "It's to get to a strong starting point sooner, so our lawyers can spend more time applying judgment, challenging the disclosure and thinking strategically about the issues that matter most to the company."
Access is also still limited to a Trusted Access cohort of U.S. firms. Most legal teams cannot try this yet, and OpenAI has not said when or at what price the API version will reach the legal technology companies already building on top of it.
Closing observation
The interesting question here isn't whether a model can sort a data room faster than an associate. It's which parts of a matter a firm is willing to hand to the first pass, and which parts it built its reputation on doing slowly, by hand, for a reason.