OpenAI introduced Astra for Law — GPT‑6 Astra wrapped with a dedicated U.S. legal search index, legal analysis/writing instructions, and a first wave of partner plugins. It is aimed at law firms and legal-tech builders who want frontier-model research and drafting without assembling the underlying legal corpus themselves.
The announcement hit #1 on Hacker News with hundreds of points and comments. That spike is why this is today’s piece: a fresh OpenAI product ship with a clear builder and legal-tech angle, not a recycled model teaser. Primary source: Introducing Astra for Law.
Important distinction (OpenAI’s own framing): Astra for Law is a configuration / product stack around GPT‑6 Astra — settings, tools, and context for professional legal work — not a separately announced new base-model weights release. In the model picker it appears as “GPT‑6 Astra Law”; the API name coming soon is gpt-6-astra-law.
Key points
- What shipped: Astra for Law = GPT‑6 Astra + legal search index + legal instructions + ecosystem plugins.
- Legal search index: U.S. case law, statutes, regulations, court rules, and administrative decisions across a corpus of more than 230 million URLs, with sources added daily. Case law draws on Free Law Project / CourtListener (OpenAI cites >99.9% of published U.S. precedential case law).
- Vendor-reported research bench: On 200 questions from Vals AI’s Legal Research Bench private validation set, at highest reasoning effort, Astra for Law passed overall correctness on 54.0% vs 38.7% for GPT‑6 Astra with web search alone (~40% relative improvement).
- Access: Initially via Trusted Access in ChatGPT and Codex for selected law firms; API “coming soon” as
gpt-6-astra-law. Named API customers: Harvey and Legora. - Controls (eligible Trusted Access firms): Zero Data Retention (ZDR) on the API; ChatGPT Enterprise usage excluded from human review by default. Governance design work with Latham & Watkins.
- Plugins / skills: 26 partner plugins (examples: iManage, Intapp, DeepJudge, Thomson Reuters HighQ / forthcoming CoCounsel connector; Relativity and Clio also named). 9 community plugins + 47 custom skills. ChatGPT for Word GA the same day.
- Base model context:
gpt-6-astra— $10 / $50 per 1M in/out, 1.05M context, 128K max out, knowledge cutoff April 2026; first Critical cybersecurity under Preparedness Framework. - Caveat: Benchmarks and demos are vendor-reported. 9to6AI has not hands-on tested Astra for Law. Legal output still needs lawyer review.
What Astra for Law is
OpenAI’s lead sentence is plain: Astra for Law combines GPT‑6 Astra with settings, tools, and context tailored for professional legal work — from researching complex cases to structuring consequential deals. Firms and legal-tech companies are meant to build products and workflows around their own expertise on top of that foundation.
The stack OpenAI describes has three practical layers:
- Model — GPT‑6 Astra (same frontier family already rolling out in ChatGPT / API / Azure / Bedrock).
- Legal search + instructions — a legal search tool over a large U.S. primary-law corpus, plus custom instructions for legal analysis and writing (e.g., distinguishing holdings from other observations, addressing adverse authority, explaining how a contract exception shifts risk).
- Connectors — partner and community plugins so ChatGPT can talk to matter systems, DMS, timekeeping, and prior-deal search instead of living as a paste box.
That is different from “a new law-only LLM with secret weights.” Secondary coverage (LawSites, SiliconANGLE, Law.com Legaltech News) matches OpenAI’s briefing language: configuration, not a new model. When we say configuration, we mean OpenAI’s product claim — we have not inspected system prompts or tool schemas ourselves.
What the legal search stack claims
OpenAI positions legal research as: find the right authority, locate the relevant passages, and judge how binding they are. The index is meant to help Astra for Law do that work and to complement licensed content from providers such as Thomson Reuters — not replace firm libraries overnight.
| Claim | OpenAI statement |
|---|---|
| Scope | U.S. case law, statutes, regulations, court rules, administrative decisions |
| Size | More than 230 million URLs; sources added daily |
| Case-law partner | Free Law Project / CourtListener; >99.9% of published U.S. precedential case law (as cited by OpenAI) |
Vals AI Legal Research Bench (OpenAI-reported, private validation set, n=200), highest reasoning effort:
| System | Overall correctness |
|---|---|
| Astra for Law | 54.0% |
| GPT‑6 Astra + web search alone | 38.7% |
OpenAI also reports ~24% more reference cases on case-law-focused questions and up to 54% more relevant target passages from correct opinions at matched reasoning effort, plus longer / more comprehensive answers on average. Treat these as vendor evals, not independent third-party replication.
Access, Trusted Access, and API timing
Today’s practical gate is not “open to every Plus user.”
- ChatGPT + Codex: Selected firms via Trusted Access; model picker label GPT‑6 Astra Law.
- API: Coming soon as
gpt-6-astra-law. Harvey and Legora are named as customers who will be able to build on it. No public API price or GA date in the Astra for Law post. - Eligible Trusted Access firms (OpenAI): Zero Data Retention on the API; ChatGPT Enterprise usage excluded from human review by default. Governance collaboration with Latham & Watkins (permissions, ethical walls, client instructions, firm oversight).
Secondary reports (LawSites / SiliconANGLE / Law.com) say the Trusted Access program is targeted at Am Law 200 firms. OpenAI’s public post says “selected law firms” without repeating “Am Law 200” in the body we verified — mark Am Law 200 as secondary / briefing-attributed, not a primary-page quote.
Firm-specific ChatGPT Enterprise builds OpenAI highlights (forward-deployed engineers + proprietary data): Sullivan & Cromwell (agreement analyzer), Ropes & Gray (deal diligence / data room), Cooley (GO Public for IPO prep). Those are named customer stories from OpenAI, not 9to6AI site visits.
Plugins and ChatGPT for Word
Same-day ecosystem drop:
- 26 partner plugins — OpenAI examples include iManage (draft → matter file), Intapp (time-entry candidates), DeepJudge (prior-deal comparison), Thomson Reuters (HighQ matter context; CoCounsel Legal connector previewed). Relativity and Clio are also named in the post.
- 9 community plugins from LegalQuants, LECG, and Skills.law, plus 47 custom skills.
- ChatGPT for Word — general availability for proofreading, suggested edits, and formatting flags inside Word.
For legal-tech builders, the interesting move is less “another chat window” and more ChatGPT as a hub that can pull matter context and push drafts back into systems of record — if your firm’s security review accepts that topology.
How this sits next to GPT-6 Astra (base model)
Astra for Law rides on GPT‑6 Astra, which OpenAI launched as its most capable broadly deployed model (rollout to Plus / Pro / Business / Enterprise, API, Azure, Bedrock; Enterprise off by default at launch per the Astra launch post). That matters for legal-tech teams building AI agents on frontier models: you inherit GPT‑6 Astra’s base limits and pricing, then add the legal search stack and instructions on top.
Verified from API model docs:
| Spec | Value |
|---|---|
| Model id | gpt-6-astra |
| Input / output | $10 / $50 per 1M tokens (cached input $1.00; long-prompt surcharges apply above 272K input) |
| Context / max out | 1,050,000 / 128,000 |
| Knowledge cutoff | April 2026 |
| Reasoning | reasoning.effort: low, medium, high, xhigh, max |
| Notable tools | Computer use, web search, file search, code interpreter, MCP, skills, … |
From OpenAI’s safety overview: GPT‑6 Astra is the first model OpenAI says reaches the Preparedness Framework’s Critical cybersecurity capability level — meaning, with the right tools and access, it can find previously unknown flaws and develop new exploit paths across many well-protected systems without step-by-step human guidance. OpenAI describes stronger cyber safeguards and notes worse chain-of-thought monitorability vs GPT‑5.6 Sol under adversarial tests. That safety profile is about the base model, not a separate Astra-for-Law safety card.
What it means for builders
- Legal-tech API surface is the real product — If you sell to firms (Harvey/Legora pattern),
gpt-6-astra-lawis the SKU to watch once it leaves “coming soon.” Until then, evaluate Trusted Access / partner plugins, not vapor pricing. For builders, the near-term surface is the forthcominggpt-6-astra-lawAPI (timing and pricing still undisclosed), in the same OpenAI ship-and-integrate pattern as OpenAI’s Realtime API. - Index + instructions ≠ free Westlaw replacement — OpenAI explicitly positions the index as complementary to licensed specialist content. Procurement still needs firm libraries, citation validation, and malpractice-aware review loops.
- 54% “correctness” is a signal, not a finish line — Even OpenAI’s own high-effort number leaves nearly half the private bench failing overall correctness. Design for retrieval + human verification, not unsupervised filing.
- Governance is part of the sale — ZDR, no-human-review defaults, ethical walls, and Latham & Watkins collaboration are how OpenAI is trying to clear Big Law security/privacy gates.
- Base Astra pricing is known; Law SKU pricing is not — Model cost on
gpt-6-astralist rates only as a rough reference until Law API pricing ships. Do not invent a Law price. - Competitive context (not this piece’s subject) — Anthropic has been pushing legal vertical positioning in the same market window; compare on access, corpus, connectors, and liability posture when that is your buying decision.
What to watch
- Public GA date and pricing for
gpt-6-astra-law. - Whether Trusted Access expands beyond the initial selected-firm set (and whether Am Law 200 targeting holds).
- Independent replications of the Vals AI Legal Research Bench numbers.
- How often partner plugins (TR / iManage / Relativity / Clio / Harvey / Legora) become production defaults vs demos.
- Interaction between Critical-cyber Astra capabilities and firm-approved agent/computer-use policies.
Sources
- Introducing Astra for Law — OpenAI (primary)
- GPT-6 Astra launch — OpenAI
- Safety overview: GPT-6 Astra — OpenAI
- GPT-6 Astra API model docs — OpenAI Developers
- Secondary (access framing / Am Law 200): LawSites, SiliconANGLE, Law.com Legaltech News
- Hacker News discussion of Astra for Law
Image credit: OpenAI
