OpenAI Astra for Law: The Small-Business AI Opportunity

What OpenAI launched on 17 September

Confirmed fact: OpenAI announced Astra for Law on 17 September 2026. The company describes it as a foundation for law firms and legal-technology businesses to build products and workflows around their own expertise.

The system combines GPT-6 Astra with instructions for legal analysis and writing and a new legal search index. OpenAI says the index searches US case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs, with new sources added daily.

OpenAI is also launching 26 partner-built plugins and 47 community plugins for legal work. Named partners include Thomson Reuters, Harvey, Legora, iManage, Clio, Intapp and Relativity. These connections are intended to bring specialist knowledge and existing firm workflows into ChatGPT while preserving the organisation's access controls.

• GPT-6 Astra configured for professional legal work
• A specialist search index focused on US legal sources
• Instructions designed for legal analysis and writing
• Plugins connecting established legal software and knowledge
• Privacy and governance controls for confidential work

Who can use it and what does it cost?

Confirmed fact: Astra for Law is initially available only to selected law firms through OpenAI's Trusted Access programme in ChatGPT and Codex. OpenAI says it will come to the API soon, where its model name will be gpt-6-astra-law.

There is no public self-service price in the launch announcement. The product page directs interested organisations to contact sales. That means a normal ChatGPT subscription should not be treated as access to Astra for Law, and small firms should not budget for it until OpenAI confirms eligibility, terms and a price for their account.

The legal search index is built around US law. A UK business must not assume the system understands the right jurisdiction simply because an answer sounds professional. The product's initial access model and source coverage make it a story to watch rather than a tool most British small businesses can deploy today.

What the performance figures really show

Confirmed fact: OpenAI tested the complete Astra for Law setup on 200 US legal-research questions from a private validation set of Vals AI's Legal Research Bench. At the highest reasoning setting, OpenAI reports that Astra for Law passed the overall correctness check on 54.0% of questions, compared with 38.7% for GPT-6 Astra using web search alone.

OpenAI describes that result as a 40% relative improvement. It also reports 24% more reference cases on case-law questions and up to 54% more relevant passages from the correct opinions when systems were compared at the same reasoning effort.

Our analysis: the improvement supports the value of specialist sources and instructions, but 54% is not permission to remove professional review. The test was reported by the product maker, used a private validation set and focused on US legal research. Real client work also includes incomplete facts, changing law, local procedure, confidentiality and judgement that a benchmark cannot fully reproduce.

Why this matters beyond law

Our analysis: Astra for Law illustrates a repeatable product pattern. Start with a capable general model, connect it to authoritative information, add instructions for one profession, fit it into existing software and require a qualified person to approve the result.

That pattern can be useful in much smaller businesses. A generic chatbot knows a little about almost everything. A controlled workflow can know which approved price list to use, which version of a procedure is current, what evidence must be attached and when a human must take over.

The defensible part is not the model alone. It is the combination of clean source material, industry knowledge, a clear sequence, tests, permissions and quality control. Those are assets a small firm can improve over time even when the underlying AI model changes.

Five practical specialist workflows to test

Choose a repetitive task where the business already owns the correct information and where a person can check the result. Avoid high-stakes decisions, regulated advice and any process that would expose sensitive data without proper controls.

A marketing agency could ground campaign checks in a client's approved brand guide. An online shop could produce product-page drafts from a verified catalogue and claims sheet. A trades business could turn survey notes into a draft quotation using an approved price book. A consultant could convert discovery notes into a structured report. A training provider could answer learner questions from the current course handbook.

These workflows do not need to act autonomously. A useful first version retrieves the correct source, produces a draft and stops for approval. That is easier to test, explain and improve than an agent allowed to make final decisions on its own.

• Brand-guide checker for agency content
• Verified-catalogue assistant for product listings
• Survey-notes-to-quotation drafting workflow
• Discovery-notes-to-client-report workflow
• Course-handbook assistant with source links

The opportunity: sell the workflow, not the chatbot

Our analysis: a freelancer or consultant can offer a specialist AI workflow audit to one narrow type of business. Map one repeated task, identify the approved sources, define the output and insert a human approval point before anything reaches a customer.

The deliverable might include a source checklist, a tested prompt or automation, an exception list, a short operating guide and a record of ten test cases. Charge for understanding and improving the process—not for pretending a generic AI wrapper is unique technology.

A strong offer makes a modest, measurable promise such as reducing the time needed to prepare a first draft. It does not promise legal accuracy, guaranteed savings or fully autonomous decisions. Start with information the client is permitted to use and document what the workflow must never do.

Risks and a practical 60-minute action plan

Astra for Law is not a replacement for a qualified solicitor, and a general AI tool should never be presented as one. Even a source-backed system can select the wrong authority, miss a change, misunderstand the facts or generate confident language around an uncertain conclusion.

Confidentiality is equally important. OpenAI says its legal API offering includes zero data retention and that ChatGPT Enterprise uses an eyes-off human-review posture by default. Those controls are relevant to eligible products and agreements; they should not be assumed for every consumer account, connector or third-party service.

For a safe test, spend 15 minutes choosing one low-risk repeated task, 15 minutes collecting the approved source documents, 15 minutes defining the required output and stop conditions, then 15 minutes testing three normal cases and three awkward ones. Record every unsupported claim and every point where a person had to intervene.

If the draft is consistently useful, test it for a week with mandatory review. If it fails on sources, exceptions or privacy, fix the process before adding automation. The lesson from Astra for Law is specialisation with evidence—not blind trust in a more impressive name.

• Pick one low-risk, repeated task
• Use only approved and current source material
• Define when the AI must stop and ask a person
• Test normal cases and difficult exceptions
• Measure correction time as well as drafting time