18 September 2026
OpenAI announced Astra for Law on 17 September 2026, a version of GPT-6 Astra configured for legal work with its own legal search index and instructions for legal analysis. On Vals AI’s Legal Research Bench it passed the correctness check on 54.0 percent of questions, against 38.7 percent for GPT-6 Astra using ordinary web search. It is available first to selected law firms through a Trusted Access programme, with API access promised but not dated.
What is Astra for Law?
It is not a new base model. Astra for Law is GPT-6 Astra, OpenAI’s flagship released on 3 September, wrapped with three things: a legal search index, a set of instructions for legal analysis and writing, and controls for confidential client work. In ChatGPT and Codex it appears in the model picker as GPT-6 Astra Law. In the API it will be called gpt-6-astra-law.
The search index is the substantive part. It covers US case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs, with new sources added daily. Free Law Project, the nonprofit behind CourtListener, supplies a case-law collection that OpenAI says covers more than 99.9 percent of published US precedential case law. The instructions then steer the model to separate a court’s holding from its side remarks, address cases that weaken an argument, and explain how a contract exception shifts risk.
Around the model, OpenAI launched 26 partner-built plugins that connect ChatGPT to tools law firms already use, including iManage, Intapp, DeepJudge, Relativity, Clio and Thomson Reuters HighQ, plus nine community plugins with 47 skills. ChatGPT for Word is generally available from today. Harvey and Legora will build on Astra for Law through the API.
Astra for Law benchmarks: how much does the legal index add?
OpenAI tested the full setup on 200 US legal research questions from the private validation set of Vals AI’s Legal Research Bench, at the highest reasoning effort for both systems. The comparison is always against its own model with web search, never against a rival.
- Overall correctness check passed: 54.0 percent for Astra for Law, 38.7 percent for GPT-6 Astra with web search, a 40 percent relative gain
- Case-law questions: Astra for Law found 24 percent more reference cases
- Audited target passages: up to 54 percent more relevant passages retrieved from the correct court opinions, at the same reasoning effort
These are OpenAI’s own evaluations, and the benchmark’s validation set is private, so nobody outside OpenAI and Vals AI can rerun them. OpenAI also shows one side-by-side example in which Claude Fable 5.1 returned a holding that had been reversed on appeal while Astra for Law returned two matching precedents. That is a single prompt, not a benchmark, and it should be read as marketing.
Who can use Astra for Law and what it costs
- Access: selected law firms through the Trusted Access Program, in ChatGPT and Codex
- API: coming soon, no date given; Harvey and Legora are named as early API builders
- Price: not published; the base GPT-6 Astra API costs $10 per million input tokens and $50 per million output tokens
- Data: Zero Data Retention on the API for eligible firms, and ChatGPT Enterprise usage excluded from human review by default
- Governance: OpenAI is working with Latham & Watkins on information permissions, ethical walls and client instructions
- Coverage: US law only, based on everything OpenAI describes in the announcement
What OpenAI is not saying about Astra for Law
The announcement never mentions a jurisdiction other than the United States. The index is built from US case law and US statutes, and the CourtListener partnership is a US collection. For a Dutch, German or UK firm the search index adds nothing, and OpenAI gives no roadmap for other legal systems. That is the single most important fact for a European reader, and it sits nowhere in the text.
There is also no comparison with the tools lawyers already pay for. Harvey, CoCounsel from Thomson Reuters and Lexis+ AI all do legal research with their own indexes, and OpenAI publishes no score for any of them on the same benchmark. The 54.0 percent figure cuts both ways: it is a large jump over web search, and it still means the system fails the correctness check on almost half of the questions. OpenAI shows a chart across reasoning-effort levels but only quotes the number at the highest setting, which is also the slowest and most expensive.
What this means
For law firms in the Netherlands and the rest of Europe this is safe to ignore as a research tool. The index does not cover your law, the price is unknown and the API has no date. What deserves attention is the pattern. A week after ChatGPT for Financial Services, OpenAI is now packaging its frontier model with a domain index, domain instructions and a plugin catalogue and calling the result a foundation for an industry. That is aimed squarely at Harvey, Legora and the other legal AI vendors, who now build on the same model their supplier sells directly to their customers. If you run or buy from a legal tech vendor, that dependency is worth a conversation this quarter.
For US firms that get Trusted Access, it is worth testing now, with one rule: treat it as a research associate whose work you check, not as an answer machine. Start with questions from your own closed matters where you know the right authorities, and compare the output with whatever tool your associates use today. A 54 percent pass rate is a real step up from a general chatbot and nowhere near good enough to file unreviewed. Google’s Gemini 3.8 Flash already cites Harvey’s Legal Agent Benchmark in its own launch, so expect the rivals to publish numbers on this same test within weeks.
For more information, visit the official announcement of Astra for Law on the OpenAI blog.
