AI
“OpenAI has launched 'Astra for Law,' a legal AI system built around GPT-6 Astra for law firms and legal technology companies, featuring a search index covering U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, and answered 54% of questions correctly on Vals AI's Legal…”
Plain restatementOn 17 September 2026 OpenAI announced Astra for Law, a configuration of its GPT-6 Astra model aimed at law firms and legal technology vendors, comprising legal instructions, tools, and a legal search index of more than 230 million URLs of US legal material. OpenAI reports that in its own testing on Vals AI's Legal Research Bench the configured system passed an overall correctness check on 54% of questions versus 38.7% for GPT-6 Astra with web search alone. OpenAI also announced 26 partner-built plugins connecting ChatGPT to law firm tools.
Distortion code this site does not recognise yet: unreleased_as_released. Not collectible until the field guide has an entry.
This post is mostly accurate. OpenAI really did launch Astra for Law on 17 September 2026, and the details check out against OpenAI's own announcement: it is a legal configuration of GPT-6 Astra, the legal search index covers US case law, statutes, regulations, court rules and administrative decisions across more than 230 million web addresses, and 26 partner plugins launched alongside it. The 54% and 38.7% figures are exact and the post correctly says they come from OpenAI. Two things are left out. First, the product is not broadly available: OpenAI says it goes initially to selected law firms through an invitation-based programme, with developer access still to come and no pricing published. Second, the test behind those numbers was run by OpenAI itself, not by the benchmark's creator Vals AI, on 200 questions from a private set at the highest reasoning setting, and a 54% pass rate also means the system failed the correctness check on nearly half the questions. No independent evaluator has yet published its own results for this system.
OpenAI [drifted from the evidence:] has launched 'Astra for Law,' a [drifted from the evidence:] legal AI system built around GPT-6 Astra [drifted from the evidence:] for law firms and legal technology [drifted from the evidence:] companies, featuring a search index [drifted from the evidence:] covering U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, [drifted from the evidence:] and answered 54% of [drifted from the evidence:] questions correctly on Vals AI's Legal Research [drifted from the evidence:] Benchmarks compared to 38.7% for GPT-6 Astra [drifted from the evidence:] using web search alone, [drifted from the evidence:] alongside the launch of 26 [drifted from the evidence:] legal plugins connecting ChatGPT [drifted from the evidence:] with law firm tools.
[added by the neutral restatement:] On 17 September 2026 OpenAI [added by the neutral restatement:] announced Astra for Law, a [added by the neutral restatement:] configuration of its GPT-6 Astra [added by the neutral restatement:] model aimed at law firms and legal technology [added by the neutral restatement:] vendors, comprising legal instructions, tools, and a [added by the neutral restatement:] legal search index [added by the neutral restatement:] of more than 230 million URLs of [added by the neutral restatement:] US legal material. OpenAI reports that in its own testing on Vals AI's Legal Research [added by the neutral restatement:] Bench the configured system passed an overall correctness check on 54% of questions versus 38.7% for GPT-6 Astra [added by the neutral restatement:] with web search alone. [added by the neutral restatement:] OpenAI also announced 26 [added by the neutral restatement:] partner-built plugins connecting ChatGPT [added by the neutral restatement:] to law firm tools.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- OpenAI did launch Astra for Law on 17 September 2026. Confirmed on OpenAI's own channel and independently by Reuters.
- It is built around GPT-6 Astra and is a configuration of that model rather than a new model.
- It is aimed at law firms and legal technology companies, with Harvey and Legora named as API partners.
- The legal search index covers US case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs. The category list in the post matches OpenAI's wording item for item.
- The figures 54% and 38.7% are exact. OpenAI's page states 54.0% versus 38.7%.
- The benchmark is correctly named in substance: Vals AI's Legal Research Bench.
- The comparison baseline is correctly stated as GPT-6 Astra using web search alone.
- 26 legal plugins connecting ChatGPT with law firm tools is correct.
- The caption attributes the benchmark result to OpenAI ("OpenAI said it answered 54%..."), which is the correct attribution.
- Unreleased as released: the post says the system has launched "for law firms and legal technology companies," with no access qualifier. OpenAI's own page says it will initially be offered to SELECTED firms through Trusted Access in ChatGPT and Codex, with API access "coming soon" and no date or pricing. A reader would reasonably conclude a law firm can obtain this today. Most cannot. This is the largest gap between the post and the source.
- Omitted qualifier: the benchmark result is reported without its harness. OpenAI's sentence specifies 200 questions, drawn from a PRIVATE VALIDATION SET, at the HIGHEST reasoning effort for both systems. Dropping "200 questions" removes the sample size, dropping "private validation set" removes the fact that this is not the public leaderboard, and dropping "highest reasoning effort" removes the compute setting. None of these reverse the finding, but together they present a bounded internal test as a plain benchmark score.
- Omitted qualifier: the absolute level is left out of the frame. Passing an overall correctness check on 54% of questions means the configured system did not pass on roughly 46%. The post reports the improvement and not the floor, which for a product sold into high-stakes legal work is the more decision-relevant number.
- Marketing as evidence, partial: OpenAI ran this evaluation itself on a third party's question set. Vals AI authored the benchmark but did not run or publish these two figures, and no Vals AI confirmation was found. The caption's "OpenAI said" prevents this from being a full distortion, but the phrasing "on Vals AI's Legal Research Benchmarks" can read as though an independent evaluator produced the result. It did not.
- Whether Vals AI has reviewed, reproduced, or endorsed OpenAI's 54.0% and 38.7% figures. No statement from Vals AI on Astra for Law was found as of 2026-09-18, one day after launch.
- How Astra for Law would score on the public Legal Research Bench leaderboard under Vals AI's own harness. Until such a run exists, the 54.0% cannot be placed against other models' published scores.
- Trial counts, grader configuration, variance across runs, and per-practice-area breakdowns behind the 54.0% figure. Not published in the announcement.
- API pricing and general availability date for gpt-6-astra-law. Not published.
- A minor discrepancy in the plugin accounting across sources. OpenAI's page and Legal IT Insider describe 9 community plugins carrying 47 custom skills, while Greg Brockman's X post says "47 community plugins." This does not affect the claim's 26 figure, which is the partner-built count and is correct in every source.
- I accessed the OpenAI announcement through search-result content extraction rather than a direct page fetch. The decisive sentences are reproduced verbatim and identically by three independent outlets, so I treat the wording as reliable, but I did not render the page.
OpenAI's own announcement page states that it is introducing Astra for Law as a foundation for law firms and legal technology companies, combining GPT-6 Astra with settings, tools, and context tailored for professional legal work, and that API customers including Harvey and Legora will be able to build on it. On the index, the announcement states that Astra for Law pairs GPT-6 Astra with instructions for legal analysis and writing, settings for thorough work, and a new Legal Search Index that searches US case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, with sources added daily. Much of the case law comes via a partnership with the Free Law Project, the nonprofit behind CourtListener, whose collection is described as covering more than 99.9% of published US precedential case law. On the benchmark, the announcement states verbatim: "we tested Astra for Law's complete setup on 200 U.S. legal research questions from the private validation set of Vals AI's Legal Research Bench," and "At the highest reasoning effort for both systems, Astra for Law passed the evaluation's overall correctness check on 54.0% of questions, compared with 38.7% for GPT-6 Astra using web search alone." OpenAI characterises this as a 40% relative improvement, and separately reports 24% more reference cases found on case-law questions and up to 54% more relevant passages retrieved from the correct opinions. On plugins, the announcement states 26 partner-built plugins for legal work in ChatGPT, from providers including Thomson Reuters, Intapp, Harvey, Legora, DeepJudge, iManage, Relativity and Clio, plus community plugins from LegalQuants, LECG and Skills.law carrying 47 custom skills. On availability, OpenAI states that Astra for Law will initially be offered to selected firms through Trusted Access in ChatGPT and Codex, with API access coming soon. It appears in the ChatGPT model picker as "GPT-6 Astra Law" and will appear in the API as gpt-6-astra-law. Reuters independently reported the launch the same day and described the product in matching terms, naming Sullivan & Cromwell, Ropes & Gray, Cooley, Latham & Watkins and Wachtell Lipton as firms OpenAI worked with.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Compact share page: ai.trueseeker.com/s/4eca81bca7c3/GSQMEuAwx-wXzKS4pGBF0phexjX
Ask this case
Answers come only from the case file above; nothing is added.
Is Astra for Law available to any law firm right now?
No. OpenAI says it will initially be offered only to selected firms through Trusted Access in ChatGPT and Codex. API access is described as coming soon, with no date or pricing published.
Where do the 54% and 38.7% accuracy figures come from?
They come from OpenAI's own testing, not from an independent run by Vals AI, which created the benchmark. OpenAI tested 200 questions from the benchmark's private validation set at the highest reasoning setting for both systems.
What do the 54% and 38.7% numbers actually mean?
OpenAI reports that Astra for Law passed an overall correctness check on 54.0% of questions, compared with 38.7% for GPT-6 Astra using web search alone. This also means Astra for Law did not pass the check on close to 46% of questions.
Has Vals AI confirmed these results?
The case file found no statement from Vals AI reviewing, reproducing, or endorsing OpenAI's figures as of one day after launch. It is also not established how Astra for Law would score on Vals AI's own public leaderboard.
What does the legal search index actually cover?
OpenAI describes it as covering US case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, with sources added daily. Much of the case law comes through a partnership with the Free Law Project.