Astra for Law Adds a Verified Case Law Index

Retrieval, Not a New Model

OpenAI introduced Astra for Law on September 17, and the framing matters more than the branding. It is not a new frontier model. It is GPT-6 Astra wrapped in domain instructions, legal tooling and — the part that does the work — a dedicated legal search index. For a profession where a confidently invented citation is a career event, the interesting question was never how fluent the model is. It is where the words come from.

• Index scale: more than 230 million URLs spanning US case law, statutes, regulations, court rules and administrative decisions
• Benchmark: 54% pass rate on Vals AI's Legal Research Bench of 200 US legal research questions, against 38.7% for GPT-6 Astra using ordinary web search
• Retrieval gains: 24% more reference cases found on case-law questions and up to 54% more relevant passages pulled from the correct opinions
• Ecosystem: 26 vendor plugins at launch including Thomson Reuters, Harvey, Legora and iManage, plus 9 community plugins and 47 custom skills

What Does the Legal Search Index Change?

General-purpose web search treats a court opinion as one more page. A legal research index treats it as a node in a citation graph with a procedural posture, a jurisdiction and a history of being affirmed, distinguished or overruled. OpenAI says the index draws on materials including CourtListener from the Free Law Project, which is a notable choice — it means a meaningful share of the corpus is openly published rather than sitting behind a proprietary subscription.

The 40% relative accuracy improvement is the headline, but the retrieval numbers underneath it are the more durable signal. Finding 24% more reference cases and 54% more relevant passages describes a system that is better at locating the right source material, not one that is better at sounding persuasive. Those are different failure modes, and the second is the one that gets lawyers into trouble.

One honest caveat: these are OpenAI's published figures against a third-party benchmark, and a 54% pass rate on legal research questions is genuine progress rather than a solved problem. Roughly half the answers still did not clear the correctness check. This is a research assistant that shortens the first pass, not a substitute for verification.

Who Can Actually Use It Right Now?

Access is deliberately narrow at launch. API customers including Harvey and Legora can build on it, and firms reach it through a Trusted Access Program that OpenAI is rolling out to large practices. API usage carries a zero data retention policy, and ChatGPT Enterprise traffic is excluded from human review by OpenAI staff — table stakes for privileged material, but worth stating plainly given what these systems would otherwise be reading.

The plugin count is the strategic tell. Twenty-six vendor integrations at launch, from the largest legal publisher through to document management and the AI-native firms, describes a platform play rather than a product launch. OpenAI is positioning Astra for Law as the layer other legal software builds on, which is the same shape as Gemini's enterprise push into law firms and a continuation of the vertical strategy we traced in GPT-6 Astra's pricing and computer-use rollout.

Why Vertical Configurations Keep Beating General Models

There is a pattern worth naming here, because it keeps repeating. The gains in Astra for Law come almost entirely from grounding rather than from raw model capability — same base model, better sources, structured tools, measurable improvement. That is cheaper to build than a new frontier model and easier to evaluate, and it suggests the next wave of professional AI products will be defined by the quality of what they can look things up in.

For legal teams the practical read is straightforward. The bottleneck in research was never drafting; it was the hours spent finding the right authority and confirming it is still good law. A system that reliably surfaces more of the relevant passages compresses that first stage, provided the output still gets checked. More on enterprise model releases and agent tooling in our AI coverage.

Sources: OpenAI — September 17, 2026; LawSites — September 17, 2026; ABA Journal — September 2026.