Astra goes
vertical, starting with law.
OpenAI has launched Astra for Law, a legal-vertical edition that beats general GPT-6 Astra on legal research tasks. Why did the company's first industry-specific product start with law? Here's what the numbers and primary sources actually say.
A search index plus tuned
instructions, not a new model
On September 17 (local time), OpenAI announced Astra for Law in its official blog post "Introducing Astra for Law". This is not a new base model — it's the same GPT-6 Astra that reached general availability on September 3, wrapped in a legal-specific search index and a set of instructions tuned for legal work.
The centerpiece is a new Legal Search Index that searches U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, updated daily. Case law comes from CourtListener, run by the nonprofit Free Law Project, which OpenAI says covers more than 99.9% of published U.S. precedent. According to reporting from The Decoder, this is paired with instructions tuned for legal analysis and drafting, plus governance settings built for confidential client work.
| GPT-6 Astra (general, with web search) | Astra for Law |
|---|---|
| Finds case law via general web search | Dedicated Legal Search Index for case law and statutes |
| Legal formatting/tone left to the prompt | Instructions pre-tuned for legal analysis and drafting |
| Overall correctness: 38.7% | Overall correctness: 54.0% |
| Broadly available via ChatGPT/API | Trusted Access (ChatGPT, Codex) → API coming soon |
What the benchmark shows
Vals AI's "Legal Research Bench," a private validation set of 200 questions, compared both systems at their highest reasoning effort.
Who can actually use it, and when
Rollout is deliberately narrow. In-house legal teams, law firms, and legal-tech vendors each face a different reality today.
Large firms (Am Law 200 scale)
Only selected firms get access, through an invitation-only "Trusted Access" program inside ChatGPT and Codex, where the model appears in the picker as "GPT-6 Astra Law." This is not a public launch.
Legal-tech vendors (Harvey, Legora)
An API is coming soon under the model ID gpt-6-astra-law. Harvey Technologies and Legora are named as early API customers planning to build it into their own products. OpenAI also lists 26 partner plugins, including Thomson Reuters, Intapp, iManage, and DeepJudge.
General in-house legal teams
Not a direct target for now. Most companies will encounter Astra for Law only indirectly, through whichever legal-tech vendor (Harvey, Legora, etc.) they already use choosing to integrate it.
Law wasn't the first place OpenAI tested
its "go vertical, industry by industry" playbook.
From one general model
to a line of industry editions
Until now, OpenAI built one general model — GPT-6 Astra — and left the industry-specific work to outside startups like Harvey and Legora. Astra for Law shows OpenAI pulling part of that work in-house. And law wasn't even the first vertical: after GPT-6 Astra reached general availability on September 3, OpenAI announced "ChatGPT for Financial Services" (built on data from LSEG, PitchBook, and Daloopa) on September 10, then Astra for Law a week later on September 17 — a roughly weekly cadence of industry editions. A similar approach reportedly preceded both of these in healthcare, suggesting this is becoming a repeatable playbook rather than a one-off product.
Why law arrived this early is a matter of inference, but it's hard to ignore that case law and statutes are unusually well-suited to this approach: sources are citable and claims are independently checkable, which is exactly what lets OpenAI publish a concrete accuracy number in the first place.
Who this actually changes things for
Business leaders
Companies evaluating or renewing legal-tech contracts now have another data point — especially if Harvey or Legora are already on the shortlist, since either could eventually run on Astra for Law under the hood. Companies with no legal-tech footprint see zero direct impact from this announcement.
Product managers
If you're building on OpenAI's API for a legal-tech product, watch for general availability of gpt-6-astra-law rather than assuming general GPT-6 Astra is your ceiling. PMs outside legal should also watch the cadence — healthcare, then finance, then law — as a leading indicator of when OpenAI might turn to their own industry next.
Recommended next steps
Legal-tech buyers should ask vendors to disclose figures against Vals AI's Legal Research Bench (54.0% overall correctness, +24% reference cases found, up to +54% relevant passages retrieved) as a baseline. PMs planning API integration should confirm the API's general-availability date rather than building around the Trusted Access preview.
This isn't unqualified good news
A 54.0% correctness rate also means Astra for Law is still wrong roughly 46% of the time on this benchmark. The comparison itself is drawn from Vals AI's private, 200-question validation set at "highest reasoning effort" for both systems — not an independent third-party audit — so it should be treated as a reference point, not a production-readiness guarantee.
The Legal Search Index covers only U.S. case law and statutes; it doesn't reach other jurisdictions or a firm's own internal work product. Access itself is narrow — an invitation-only "Trusted Access" preview aimed at Am Law 200-scale firms, not a general release most companies can try today. There's also a structural wrinkle worth watching: Harvey and Legora, named as OpenAI's own early API customers, have spent years building legal-AI products on top of OpenAI's models as customers. OpenAI now competing in the same space under its own brand is a relationship worth monitoring for channel conflict.