{"slug": "astra-for-law-is-not-a-new-model-what-openai-added", "title": "Astra for Law Is Not a New Model: What OpenAI Added", "summary": "OpenAI introduced Astra for Law on September 17, 2026, a legal configuration built on its existing GPT-6 Astra flagship model rather than a new model, combining a search index over US law, custom instructions, a higher effort setting, a Trusted Access Program gate for selected law firms, and zero data retention with exclusion from human review by default. OpenAI's own benchmark scored Astra for Law at 54.0% versus 38.7% for GPT-6 Astra with web search alone on 200 questions from a private validation set run by OpenAI, with no comparison against Westlaw, Lexis, Harvey or CoCounsel. API availability is promised \"coming soon\" with no date and no published price, while 26 partner plugins and a Word integration shipped the same day through ChatGPT's plugin ecosystem.", "body_md": "On September 17, 2026 OpenAI introduced Astra for Law. Read the announcement carefully and it is not a new model. OpenAI's own words are that it combines GPT-6 Astra, the company's current flagship, with settings, tools and context for professional legal work. The model is the same one any API customer can already buy. What is new sits around it: a search index over US law, a set of written instructions, a higher effort setting, an access gate for law firms, and confidentiality terms that most business customers do not get by default.\n\nThat distinction matters beyond law. Every regulated vertical is about to be offered an \"industry version\" of a frontier model, and this launch shows what one contains. This post takes the package apart, reads the benchmark the way a buyer should, and sets out the pattern a firm in any sector can copy without waiting for a lab to sell it to them. Every fact below is from [OpenAI's announcement](https://openai.com/index/astra-for-law/), read on September 22, 2026. No price is published, and no date is given for API availability.\n\n1. 01It is a configuration, not a model.GPT-6 Astra plus a legal search index, custom instructions and thorough-work settings. It appears in the model picker as 'GPT-6 Astra Law' and will have the API id gpt-6-astra-law, but the underlying model is unchanged.\n2. 02The benchmark compares OpenAI with OpenAI.54.0% against 38.7% is Astra for Law against GPT-6 Astra with web search alone, on 200 questions from a private validation set, run by OpenAI. No comparison with Westlaw, Lexis, Harvey or CoCounsel exists on the page.\n3. 03The governance terms are the part worth copying now.Zero data retention on the API and exclusion from human review by default are what a firm should ask any AI vendor to match, whichever model sits underneath.\n4. 04Access is gated, and the gate is not defined.Selected law firms get it through a Trusted Access Program in ChatGPT and Codex. OpenAI does not say how a firm qualifies. Anthropic's legal package predates this one, so it is not a first.\n\n## 01 — The launchWhat it is, and who can get it\n\nOpenAI calls Astra for Law \"a new foundation for law firms and legal technology companies\". Underneath that phrase the announcement is precise. The intelligence is GPT-6 Astra. The legal part is a search index the model can call, instructions that shape how it analyses and writes, and settings for thorough work. OpenAI says it will carry these legal capabilities forward to future models, which confirms the reading: the legal layer is separable from the model it currently sits on.\n\nAvailability comes in three tiers. Selected law firms get it first, through a Trusted Access Program, inside ChatGPT and inside Codex, OpenAI's coding agent. API customers, with Harvey and Legora named, will be able to build on it \"coming soon\", with no date. And the 26 partner plugins and the Word integration shipped the same day through ChatGPT's plugin ecosystem rather than through the gated legal configuration; OpenAI does not publish which ChatGPT plans they require. We covered the base model's price and published benchmarks in [our GPT-6 Astra guide](https://www.digitalapplied.com/blog/gpt-6-astra-price-benchmarks-guide); nothing there changes.\n\nIt combines GPT‑6 Astra, our latest and most powerful model, with settings, tools, and context tailored for professional legal work.OpenAI, 'Introducing Astra for Law', September 17, 2026\n\n## 02 — The partsThe *eight parts*, and who gets each one\n\nThe table splits the announcement into its components. The \"who gets it\" column matters because the parts are sold through different doors: the configuration through a gated programme, the plugins through ChatGPT's business plans, and the firm-built tools through direct engineering work with OpenAI.\n\n| Components as described on OpenAI's announcement of September 17, 2026, read September 22, 2026. |  |  | \n|---|---|---|\n| Component | What it is | Who gets it | \n|---|---|---|\n| The model | GPT-6 Astra, OpenAI's current flagship, unchanged. The page describes everything else as settings, tools and context added around it. | Trusted Access firms in ChatGPT and Codex; API customers 'coming soon' | \n| Legal search index | A tool the model can call that searches US case law, statutes, regulations, court rules and administrative decisions across what OpenAI says is more than 230 million URLs, with case law from the nonprofit Free Law Project's CourtListener collection. | Part of the Astra for Law configuration | \n| Written instructions | Custom instructions for legal analysis and writing: applying research to the client's facts, distinguishing a holding from other remarks, addressing cases that cut against the argument. | Part of the configuration | \n| Effort settings | Settings for thorough work. The benchmark figures are at the highest reasoning effort. | Part of the configuration | \n| Access gate | A Trusted Access Program for eligible law firms covering lawyers and the people working under their supervision. | Selected law firms; OpenAI's page does not define eligibility | \n| Governance terms | Zero Data Retention on the API and ChatGPT Enterprise usage excluded from human review by default. | Eligible firms in the programme | \n| Plugins | 26 partner-built plugins connecting ChatGPT to tools such as Relativity, Clio, iManage, Intapp and DeepJudge, plus 9 community plugins with 47 custom skills. ChatGPT for Word became generally available the same day. | ChatGPT plugin users; OpenAI's page does not state the plan required | \n| Firm-built tools | Custom interfaces and integrations built with OpenAI's forward-deployed engineers on ChatGPT Enterprise, such as an agreement analyzer and a deal-diligence system. | Selected firms working directly with OpenAI | \n\n## 03 — The numbersThe benchmark, and what it does not compare\n\nOpenAI tested the full configuration on 200 US legal research questions drawn from the private validation set of Vals AI's Legal Research Bench. Vals AI is a third-party benchmark publisher; the private set is one the public cannot inspect. Both systems ran at the highest reasoning effort. The four figures below are OpenAI's own, and each one compares Astra for Law with GPT-6 Astra using web search alone.\n\n- Overall correctness check passedAstra for Law vs GPT-6 Astra with web search; 200 questions; highest effort; vendor-run\n- 54.0% vs 38.7%\n- Relative improvement, as OpenAI states it15.3 percentage points on the same 200 questions\n- 40%\n- More reference cases found on case-law questionshighest effort; vendor-run\n- +24%\n- More relevant passages retrieved from the correct opinionsaudited set of target passages; same effort for both; 'up to'; vendor-run\n- up to +54%\n\nThree things the page does not do. It does not compare Astra for Law with any legal research product: no Westlaw, no Lexis, no Harvey, no CoCounsel. It does not compare with any other lab's model on the benchmark; the one cross-vendor example is a single prompt where OpenAI reports that a competing model returned a reversed holding, which is an anecdote rather than a measurement. And it does not publish the baseline for the model without any search at all, so a reader cannot separate the value of the index from the value of the instructions. The honest summary is that the legal layer makes OpenAI's own model materially better at OpenAI's chosen research test, by OpenAI's measurement.\n\n## 04 — GovernanceThe governance terms to ask every vendor for\n\nThe most transferable part of the launch is two sentences about data. For eligible firms, API usage runs under Zero Data Retention, meaning OpenAI does not keep prompts and outputs after serving them. And ChatGPT Enterprise usage is excluded from human review by default, meaning no OpenAI staff read a firm's conversations unless the firm changes that setting. OpenAI also names Latham & Watkins as its partner on information permissions, ethical walls, client instructions and firm oversight, which are the controls a firm needs before a model can see more than one client's matters.\n\nNone of those terms depend on the legal index. They are contract and configuration terms, and a firm evaluating any AI vendor, on any model, can ask for the same four things: no retention of prompts, no human review by default, per-matter permission boundaries, and an oversight log the firm controls. Our [legal-team playbook for contract review agents](https://www.digitalapplied.com/blog/agentic-ai-legal-team-playbook-contract-review-agents-2026) covers how to run that evaluation.\n\nOpenAI uses the same phrase, Trusted Access, for its programmes that gate cyber-capable behaviour in GPT-6 Astra, which we covered in [our post on Astra's critical cyber threshold](https://www.digitalapplied.com/blog/openai-astra-critical-cyber-threshold-agent-controls). The legal programme gates by profession rather than by risk category, but the mechanism is the same. The page does not publish the eligibility criteria, the application route beyond \"contact OpenAI\", or a timeline for wider access.\n\n## 05 — The patternThe pattern any vertical can copy\n\nStrip the legal branding and Astra for Law is a recipe with five ingredients: a curated source index the model searches instead of the open web, written instructions for the profession's way of reasoning, an effort setting tuned for thoroughness over speed, an access gate that limits who can use it, and plugins into the tools the profession already runs. A finance team, a clinical team or an engineering consultancy can build the first three today on any frontier model, because they are retrieval, prompting and a parameter. The gate and the plugins are what a lab or a specialist vendor adds.\n\nThe decision for a buyer is therefore not \"which model\" but \"who assembles the package\". The router below is how we would frame it for a client.\n\nWhichever route a firm takes, the mistake to avoid is treating an industry version as a different model that needs re-evaluating from scratch. Evaluate the index, the instructions and the terms. The model underneath is the one you already know. If you want help assembling a configuration like this for your own sector, our [AI transformation service](https://www.digitalapplied.com/services/ai-transformation) builds exactly that layer.\n\n## 06 — The marketWhat legal-tech vendors now face\n\nTwo things at once. The model layer they built on has become a legal model layer, which lowers the cost of the research and drafting parts of their products and lets OpenAI sell those parts directly to firms through ChatGPT. And the same launch names them as partners, with Harvey and Legora building on the API and Thomson Reuters bringing matter context and a connector preview into ChatGPT, which means their value has to move to the parts OpenAI says it will not build: the workflow, the licensed content, the matter file, the review. Anthropic's legal package, which we analysed in [our post on the Claude legal plugin](https://www.digitalapplied.com/blog/anthropic-claude-legal-plugin-market-impact), put the same pressure on the same vendors earlier; this launch adds a search index and a benchmark to it. Our census of every lab's industry package, [across law, finance and health](https://www.digitalapplied.com/blog/ai-labs-industry-versions-law-finance-health-what-each-adds), shows how consistent the recipe has become.\n\n## 07 — Next stepThe model is old news; the package is the news\n\n### Ask what sits around the model before you ask which model\n\nIf a vendor offers you an industry version of anything, ask for the five ingredients by name: which sources are indexed, what the instructions say, what effort setting is used, who is allowed in, and which of your tools it connects to. Then ask for the two data terms OpenAI gave law firms, in writing. A package that cannot answer those questions is a model with a new label.", "url": "https://wpnews.pro/news/astra-for-law-is-not-a-new-model-what-openai-added", "canonical_source": "https://www.digitalapplied.com/blog/astra-for-law-not-a-new-model-what-openai-added", "published_at": "2026-09-19 00:00:00+00:00", "updated_at": "2026-09-22 09:24:09.359453+00:00", "lang": "en", "topics": ["ai-products", "artificial-intelligence", "large-language-models", "ai-policy"], "entities": ["OpenAI", "Astra for Law", "GPT-6 Astra", "ChatGPT", "Codex", "Harvey", "Legora", "Trusted Access Program"], "alternates": {"html": "https://wpnews.pro/news/astra-for-law-is-not-a-new-model-what-openai-added", "markdown": "https://wpnews.pro/news/astra-for-law-is-not-a-new-model-what-openai-added.md", "text": "https://wpnews.pro/news/astra-for-law-is-not-a-new-model-what-openai-added.txt", "jsonld": "https://wpnews.pro/news/astra-for-law-is-not-a-new-model-what-openai-added.jsonld"}}