OpenAI introduced Astra for Law on September 17, 2026 — not a new model, but GPT-6 Astra configured specifically for legal work: a 230-million-source legal search index, instructions tuned for legal analysis and writing, and a Trusted Access rollout to select law firms with data-privacy guarantees built in. OpenAI claims a 40% relative improvement over plain web-search Astra on a private legal research benchmark, and partners including Harvey, Sullivan & Cromwell, Ropes & Gray, and Cooley are already building on it. Here's what actually shipped, what the benchmark numbers do and don't establish, and how to read OpenAI's own comparison against Claude.
TL;DR
| Question | Answer |
|---|---|
| What is it? | GPT-6 Astra plus a legal search index (230M+ sources) plus legal-analysis instructions — a vertical configuration, not a new model |
| When did it launch? | September 17, 2026 |
| Who has access? | Selected law firms via Trusted Access in ChatGPT and Codex today; general API access "coming soon" |
| How much better is it? | 54.0% vs. 38.7% correctness on Vals AI's Legal Research Bench validation set (OpenAI's own benchmark, not independently audited) |
| What's the data source? | 230M+ URLs of case law, statutes, regulations, and administrative decisions, including Free Law Project's CourtListener case-law corpus |
| What are the privacy controls? | Zero Data Retention on the API for Trusted Access firms; ChatGPT Enterprise usage excluded from human review by default |
| Who's building on it? | Harvey, Legora as API partners; Sullivan & Cromwell, Ropes & Gray, and Cooley built custom internal tools; 26 partner + 47 community plugins ship at launch |
What Astra for Law actually is
It's important to be precise about what changed here, because the framing matters. Astra for Law is not a new foundation model — it's GPT-6 Astra, OpenAI's latest and most capable model, wrapped in three things specific to legal practice: a purpose-built retrieval index, a set of instructions for legal analysis and writing, and firm-specific customization built by OpenAI's forward-deployed engineers working directly with law firms.
The retrieval layer is the most concrete piece. OpenAI's Legal Search Index covers more than 230 million URLs of U.S. case law, statutes, regulations, court rules, and administrative decisions, with sources added daily. A partnership with the Free Law Project, the nonprofit behind CourtListener, brings in case-law coverage OpenAI describes as more than 99.9% of published U.S. precedential case law. OpenAI is explicit that this complements rather than replaces licensed products firms already use, name-checking Thomson Reuters specifically — and Thomson Reuters itself is a launch partner, bringing HighQ matter context into ChatGPT and previewing a CoCounsel Legal connector.
The instructions layer is less visible from the outside but is where the actual legal-reasoning behavior lives: guidance for distinguishing a court's holding from its other observations, addressing cases that weaken an argument rather than ignoring them, and explaining how a contract exception shifts risk between parties. This is the same general pattern explainx.ai has covered in Claude's own knowledge-work plugin approach — taking a frontier general model and layering domain-specific retrieval and instructions on top, rather than training a separate specialized model from scratch.
The benchmark: 54.0% vs. 38.7%
OpenAI tested Astra for Law's complete setup on 200 U.S. legal research questions drawn from the private validation set of Vals AI's Legal Research Bench, a benchmark that measures how well a model finds relevant sources and passages and how well its research answers meet defined evaluation criteria. At the highest reasoning effort setting for both systems, Astra for Law passed the benchmark's overall correctness check on 54.0% of questions, compared with 38.7% for plain GPT-6 Astra using web search alone — a relative improvement OpenAI states as 40%.
Two supporting numbers worth noting: on case-law-focused questions, Astra for Law found 24% more reference cases than web-search Astra at the same reasoning effort, and retrieved up to 54% more relevant passages from correct court opinions on an audited subset of target passages. All of this is worth reading the way explainx.ai reads any vendor-published benchmark claim: it's a real, named, third-party benchmark rather than an invented internal metric, which is better than nothing — but it's OpenAI's own testing run, on a benchmark OpenAI selected, without independent reproduction disclosed alongside the launch.
The Claude comparison, and why it deserves more scrutiny than the legal numbers
OpenAI's launch post includes two worked examples — one litigation memo, one transactional memo — comparing Astra for Law's output against Claude Fable 5.1's output on the same prompt. In the litigation example, OpenAI's write-up states that Claude returned a precedent whose holding had been reversed on appeal; in the transactional example, it states Claude reported finding no matching case at all, while Astra for Law returned two.
Read this the way explainx.ai reads any single-vendor comparison of a competitor's model, which is with real caution: it's a hand-picked pair of examples selected and presented by the company launching the product being favorably compared, with no disclosed methodology for how those two prompts were chosen out of however many were tried, no independent judge, and no raw transcripts published for a reader to check the reversed-on-appeal claim directly. That doesn't make the claim false — citing an overturned case is a real, well-documented failure mode across LLMs, not just Claude, and one explainx.ai has covered directly in AI hallucinated legal citations getting lawyers sanctioned. It does mean this specific comparison belongs in the category of marketing evidence, not benchmark evidence, and shouldn't be weighted the same as the Vals AI numbers above it.
Trusted Access, privacy, and what firms are actually building
Legal work runs on confidentiality, so the access model matters as much as the model itself. OpenAI's Trusted Access Program gives eligible firms Zero Data Retention (ZDR) on the API, and usage of ChatGPT Enterprise is excluded from human review by default. OpenAI is also working with Latham & Watkins specifically on information permissions, ethical walls, client instructions, and firm oversight — governance infrastructure that matters more to a law firm's actual adoption decision than the benchmark score.
On the product side, OpenAI's forward-deployed engineers worked directly with three named firms to build custom tools on top of Astra for Law: Sullivan & Cromwell built an agreement analyzer that uses the firm's own negotiating playbooks to flag risks and draft redlines; Ropes & Gray built a data-room diligence system that traces findings back to source documents; Cooley built "GO Public," a tool for IPO filing preparation that propagates changes across a filing as a deal evolves. Harvey and Legora are named as API partners building their own products on top of Astra for Law rather than using OpenAI's own interface.
At launch, Astra for Law ships with 26 partner-built plugins (Thomson Reuters, Harvey, Legora, iManage, Intapp, DeepJudge, and others) and 47 community-built plugins and skills from practitioners at firms including LegalQuants, LECG, and Skills.law. OpenAI is also making ChatGPT for Word generally available alongside this launch, for proofreading and inline edit suggestions in the drafting tool lawyers already use.
Honest limitations
- Trusted Access is selective, not general availability — most firms don't have this yet, and general API access is only described as "coming soon" with no date given.
- The 54.0%/38.7% benchmark numbers are OpenAI's own reported results on a private validation set, without an independent third-party reproduction disclosed at launch.
- The Claude comparison is marketing, not a controlled benchmark — two hand-selected examples from the company launching the competing product, with no published methodology or raw transcripts.
- This is a configuration of GPT-6 Astra, not a standalone legal model — its ceiling is tied to the underlying model's general reasoning quality, with retrieval and instructions layered on top rather than legal reasoning trained in from scratch.
What this means for builders
Astra for Law is a template worth watching regardless of whether you work in legal AI specifically: take a frontier general-purpose model, attach a domain-specific retrieval index that's actually current (sources added daily, not a static snapshot), write instructions that encode the domain's actual reasoning patterns rather than generic prompting, and layer in the access controls the vertical actually needs before worrying about raw capability. That's a more replicable playbook for a vertical AI product than training a specialized model, and it's the same shape Anthropic has used for its own vertical Claude offerings in adjacent knowledge-work domains.
Related on explainx.ai
- AI hallucinated legal citations are getting lawyers sanctioned
- AI and the law: using AI for legal help and contracts
- How to read AI benchmark claims critically
- AI benchmark claims: a fact-check framework
- Claude's knowledge-work plugins guide
- GPT-6 Astra vs. Claude Fable 5.1: which model wins where
- GPT-6 Astra's launch benchmarks and pricing
- Official source: OpenAI — Introducing Astra for Law
This post is sourced to OpenAI's own September 17, 2026 launch materials. Benchmark figures, the Claude comparison, and partner statements are OpenAI's self-reported claims; no independent third-party audit of the Legal Research Bench results was available at time of writing.