California Law Firm SEO is a ByteZero property that plans and builds organic search and AI-citation programs for law firms in California — Los Angeles, San Diego, San Francisco, Sacramento, Fresno, Bakersfield, Stockton, San Bernardino and Chula Vista among them. On September 17, 2026, OpenAI introduced Astra for Law, a configuration of its GPT‑6 Astra model with a legal search index behind it. Most of the coverage has led on one figure: a 40% relative improvement in research correctness. The figure underneath it is the one that should shape how a California firm actually uses the thing. On OpenAI's own benchmark, at the highest reasoning effort, Astra for Law passed the overall correctness check on 54.0% of 200 legal research questions. That is the improved score.
Key Takeaways
- It is a configuration, not a new model. GPT‑6 Astra plus a legal search index, custom instructions for legal analysis and writing, and adjusted settings. It appears in the model picker as "GPT‑6 Astra Law" and in the API as gpt-6-astra-law.
- The headline number is 54.0%, not 94%. Against 38.7% for GPT‑6 Astra on web search alone, on 200 questions from the private validation set of Vals AI's Legal Research Bench, at the highest reasoning effort for both systems. A 40% relative improvement on a base that is still under half.
- You almost certainly cannot get it today. It is offered first to selected firms through a Trusted Access Program in ChatGPT and Codex, aimed at the Am Law 200 and legal technology vendors. API access is coming.
- The verification duty moves; it does not retire. COPRAC's proposed comment to Rule 3.3 reaches authority that is fabricated, misstated or taken out of context. A retrieval index helps with the first of those three and does nothing about the other two.
- Time saved is not time billed. The State Bar's Practical Guidance permits charging for time actually spent on prompting and review, and says a lawyer must not charge the client for the time the tool saved. For an hourly practice that is a revenue decision, not a footnote.
What OpenAI actually shipped on September 17
The announcement is unusually specific, which makes it easy to separate what was built from what was implied. The short version is that OpenAI did not release a legal model. It released a legal configuration of a general model, plus a research index for it to call, plus an ecosystem to plug into.
| Element | What OpenAI states |
|---|---|
| The model | Not new. GPT‑6 Astra with settings, tools and context tailored for professional legal work |
| The index | US case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs, with sources added daily |
| Case-law coverage | Through the Free Law Project, CourtListener's collection covering more than 99.9% of published US precedential case law |
| The benchmark | 54.0% overall correctness against 38.7% for GPT‑6 Astra on web search alone, on 200 questions from the private validation set of Vals AI's Legal Research Bench, at the highest reasoning effort for both |
| Retrieval gains | 24% more reference cases on case-law questions; up to 54% more relevant passages from the correct court opinions, compared at the same reasoning effort |
| Access | Trusted Access in ChatGPT and Codex for selected firms; API to follow |
| Where you would see it | "GPT‑6 Astra Law" in the model picker, gpt-6-astra-law in the API |
| Ecosystem | 26 partner plugins including Thomson Reuters, iManage, Intapp, DeepJudge, Relativity and Clio; 9 community plugins carrying 47 skills; ChatGPT for Word generally available |
Two details in that table deserve more attention than they are getting. The first is that OpenAI positions the index as a complement to licensed research products rather than a replacement, naming Thomson Reuters directly. A firm reading the launch as permission to cancel its research subscription is reading past the vendor's own framing.
The second is what the flagship firms actually built. Sullivan & Cromwell described an agreement analyzer carrying its own negotiating playbooks; Ropes & Gray a deal diligence system; Cooley a capital-markets tool for IPO preparation. All three were built with OpenAI's forward-deployed engineers on top of ChatGPT Enterprise, against each firm's proprietary data. That is the real product at the top of the market: bespoke engineering wrapped around a model, using knowledge the firm already owned. It is not something a five-attorney practice in Van Nuys buys off a pricing page, and pretending otherwise is how firms end up disappointed by tools that were never aimed at them.
The number worth holding on to is 54.0%
The 40% relative improvement is real, and on a research benchmark it is a large move. But relative improvements describe a trajectory, and a lawyer has to work in the absolute. At the highest reasoning effort, on questions drawn from a benchmark the vendor selected, roughly 46 of every 100 answers did not clear the evaluation's overall correctness check.
Read that carefully, because it cuts in both directions. The Legal Research Bench's correctness check is a strict composite: it measures source retrieval, passage retrieval and answer quality together. A failed item is not necessarily an invented case. It can be an answer that was correct but incomplete, or one that missed an authority a careful associate would have found. So 54.0% understates how useful the tool is on an ordinary Tuesday.
It also means the failure modes are the quiet ones. A hallucinated citation announces itself. A missing authority does not. The defect that costs a California firm a motion is far more likely to be the case nobody surfaced than the case nobody checked, and better retrieval improves the odds without ever closing the gap.
A research tool that is right 54% of the time is not a research tool you delegate to. It is a research tool you supervise.
OpenAI's own launch material includes a worked comparison in which a competing frontier model returned a holding that had been reversed on appeal. Treat the comparison itself as advocacy — it is a vendor characterising a rival in its own announcement. But treat the failure mode as instructive, because it is the one that gets lawyers sanctioned: a real case, real text, and a status nobody checked. No index of 230 million URLs fixes that, because subsequent history is a question about what happened to a decision after it was published.
You almost certainly cannot get it yet, and that is fine
Astra for Law is offered first through a Trusted Access Program for eligible law firms, in ChatGPT and Codex, with the API following. The stated target is the Am Law 200 and legal technology vendors. For eligible firms the offering includes Zero Data Retention on the API, and ChatGPT Enterprise usage excluded from human review by default. OpenAI also named Latham & Watkins as a collaborator on information permissions, ethical walls, client instructions and firm oversight.
If your firm is not in that set — and the overwhelming majority of California firms are not — three things follow.
Know the tell
If your model picker does not say "GPT‑6 Astra Law", you are not using Astra for Law. You are using GPT‑6 Astra with web search, which is the 38.7% column in OpenAI's own table, not the 54.0% one. This matters more than it sounds. The most common way a lawyer will get burned in the next six months is by reading a launch announcement, opening a consumer chat window, and assuming the capability described is the capability in front of them.
Confidentiality attaches to the plan, not to the model
The Zero Data Retention and human-review terms described in the announcement attach to eligible firms' API and ChatGPT Enterprise usage. A personal account has neither. Rule 1.6 of the California Rules of Professional Conduct and Business and Professions Code section 6068(e)(1) do not soften because the tool got better. If the firm is going to put client facts anywhere, it buys the plan whose terms it can defend and writes down which plan that is.
Waiting costs nothing
Everything that actually needs doing in the next quarter — the written policy, the verification step, the client disclosure sentence, the billing rule — is free, is vendor-neutral, and is close to what the State Bar is proposing to require anyway. A firm that spends this quarter on governance rather than on a waitlist will be in a better position when access widens than one that did the reverse.
What California already requires, and what it is proposing to require
California has had written guidance on this since before most firms were paying attention. The State Bar's Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law was issued on November 16, 2023, and the Board of Trustees approved revisions to it on May 14, 2026. It is framed as guiding principles rather than best practices, and it organises the obligations around confidentiality, competence, supervision, billing, candor and fairness.
The bigger development is that the principles are being proposed for the Rules themselves. On August 22, 2025, the State Bar received a letter from the California Supreme Court directing it to consider incorporating the 2023 guidance into the Rules of Professional Conduct, and to consider further guidance in light of agentic tools that can carry out workflows without human prompting. COPRAC approved proposed amendments to six rules at its March 13, 2026 meeting, and the 45-day public comment period closed on May 4, 2026.
These are proposals, not rules. A rule change in California runs through the Board of Trustees and then the California Supreme Court, and nothing below is in force on the strength of the proposal alone. Check the current status before you rely on any of it. What makes the proposals worth reading now is that they tell you, in the regulator's own words, where the duty is understood to sit.
| Rule | What the proposed comment adds | What it asks of a firm |
|---|---|---|
| 1.1 Competence | Names AI as relevant technology, and adds that a lawyer must independently review, verify and exercise professional judgment regarding any output the technology generates | A named human reviews every output that reaches a client or a court |
| 1.4 Communication | Where the use of technology presents significant risk or materially affects the scope, cost, manner or decision-making process of the representation, the lawyer communicates enough for the client to make informed decisions, through the life of the matter | Engagement-letter language, plus a trigger to re-disclose when the tooling changes |
| 1.6 Confidentiality | Defines "reveal" to include exposing confidential information to technological systems where that exposure creates a material risk of use inconsistent with the duty | An approved-tools list, and a written rule about what may be pasted where |
| 3.3 Candor | The duty of candor includes verifying the accuracy and existence of cited authorities, including that none is fabricated, misstated or taken out of context, before submission | A cite-check step that reads the case, rather than confirming it exists |
| 5.1 Managerial lawyers | Reasonable efforts to establish internal policies and procedures governing the use of AI | A written policy owned by a named partner |
| 5.3 Nonlawyer assistants | Instruction and supervision extends to the use of technology, including AI, in providing legal services | Documented paralegal and staff training |
Rule 3.3 is where Astra for Law and California actually meet
The proposed comment to Rule 3.3 does not stop at fabricated authority. It reaches authority that is misstated or taken out of context. That distinction is the whole argument of this article in one line.
A retrieval index over 230 million URLs makes fabrication materially less likely, because the model is pulling text out of real documents rather than reconstructing something that reads like one. It does nothing at all about a real case characterised a shade too favourably, a holding stretched past what the court decided, or a passage quoted without the sentence that qualifies it. Those are the second and third failure modes the proposed comment names, and they are precisely the ones that survive better retrieval.
So the firm that reads "230 million URLs" and relaxes its cite-check has moved in exactly the wrong direction. The right move is the opposite: as fabricated citations get rarer, the remaining errors get harder to spot, and the review step has to get sharper rather than lighter.
The billing question a smaller firm hits first
The Practical Guidance is unusually direct on fees. A lawyer may use generative AI to create work product more efficiently, and may charge for the time actually spent — refining inputs and prompts, reviewing outputs. A lawyer must not charge the client for the time the tool saved.
For an hourly practice that is an operational decision before it is an ethical one. If a research memo that took six hours takes two, the firm bills two, plus the prompting and review time. The other hours do not reappear on the invoice. A firm that adopts these tools aggressively on hourly work is choosing to compress a revenue line, and that is a decision a managing partner should make deliberately rather than discover in a realisation report nine months later.
A contingency or flat-fee practice reaches the value question on a completely different timeline, because there the saved hours convert directly into capacity. Which model your firm runs is therefore a better predictor of how much this launch matters to you than your practice area is.
The part of this that reaches your phone
Astra for Law's index is primary law. It is not a directory of attorneys, and nothing in the announcement changes how a person in California finds a lawyer.
What the announcement does change is the surface. Twenty-six plugins, ChatGPT for Word going generally available, and three of the most recognisable firms in the country describing custom builds amounts to the profession being told, loudly, that this is where legal work now happens. Consumer habit follows professional habit, and consumer habit is what decides whether someone in Los Angeles with a wage claim types their question into a search box or into an assistant.
That is a different optimisation problem from ranking, and it is the one we spend most of our time on. What actually gets quoted covers the sentence-level test a page has to pass before a model can lift anything from it, and GEO for California attorneys covers how the program is built. The short version: an assistant cannot cite a page that never makes a self-contained, checkable claim, and most law firm copy never makes one. The firms that get quoted in 2027 are writing those sentences now.
The advertising rules follow you onto that surface too. Anything a model can lift from your site is a communication about your services, and Chapter 7 of the Rules of Professional Conduct does not stop applying because a machine did the quoting. We covered the current position in California attorney advertising rules.
What a California firm should actually do this quarter
- Write the policy. One page. Approved tools, forbidden inputs, the review step, and the partner who owns it. Proposed Rule 5.1 asks managerial lawyers for reasonable efforts to establish exactly this, and it is the cheapest item on the list.
- Fix the cite-check, not the tool. Every authority that reaches a filing gets read, not confirmed. Existence is not the test. Proposed Rule 3.3 names misstatement and context alongside fabrication.
- Separate the accounts. Personal chat accounts are not a place for client facts. Decide which plan the firm uses, read its retention terms, and write the answer down where staff can find it.
- Draft the client sentence once. Proposed Rule 1.4 asks for enough information for the client to decide, where the technology materially affects scope, cost, manner or decision-making. Put it in the engagement letter and revisit it whenever the tooling changes.
- Settle the billing rule before the first invoice. Time actually spent, never time saved. Tell the billing partner and the timekeepers on the same day.
- Check where you are cited, not only where you rank. Ranking and citation are now two different scoreboards. Our free AI visibility audit shows which assistants currently name your firm on the questions your clients ask, and which name a competitor instead.
Frequently asked questions
What is Astra for Law?
Astra for Law is a configuration of OpenAI's GPT‑6 Astra model, introduced on September 17, 2026, built for professional legal work. It is not a new model. It combines GPT‑6 Astra with a legal search index covering US case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs, plus custom instructions for legal analysis and writing. It appears in the model picker as "GPT‑6 Astra Law" and in the API as gpt-6-astra-law.
How accurate is Astra for Law?
On OpenAI's published test, Astra for Law passed the overall correctness check on 54.0% of 200 US legal research questions from the private validation set of Vals AI's Legal Research Bench, against 38.7% for GPT‑6 Astra using web search alone, at the highest reasoning effort for both systems. That is a 40% relative improvement on a base still under half. The benchmark's correctness check is a strict composite of source retrieval, passage retrieval and answer quality, so a failed item is not necessarily an invented case — it may be an incomplete answer or a missed authority.
Can a small California law firm get Astra for Law?
Not readily, as of the launch. OpenAI is offering it first to selected law firms through a Trusted Access Program in ChatGPT and Codex, with the stated target being the Am Law 200 and legal technology vendors, and API access following. If your model picker does not show "GPT‑6 Astra Law", you are using the general GPT‑6 Astra model with web search, which is the lower column in OpenAI's own benchmark table.
Does better AI legal research remove the duty to check citations?
No, and in California the proposed rules point the other way. COPRAC's proposed comment to Rule 3.3 would confirm that the duty of candor includes verifying the accuracy and existence of cited authorities, including that none is fabricated, misstated or taken out of context, before submission to a tribunal. A retrieval index reduces fabrication. It does nothing about a real case whose holding is overstated, whose context is dropped, or whose subsequent history nobody checked.
Can a California lawyer bill the time an AI tool saves?
No. The State Bar's Practical Guidance permits charging for the time actually spent on the work product, including refining inputs and prompts and reviewing outputs, and states that a lawyer must not charge the client for the time saved by using generative AI. For firms on hourly billing, adopting these tools is therefore a deliberate decision to compress a revenue line rather than a free efficiency gain.
Is it safe to put client information into ChatGPT?
It depends entirely on the plan and its terms, not on the model. OpenAI's announcement describes Zero Data Retention on the API and exclusion of ChatGPT Enterprise usage from human review by default, for firms eligible under the Trusted Access Program. A personal consumer account carries neither. Rule 1.6 of the California Rules of Professional Conduct and Business and Professions Code section 6068(e)(1) still govern, and COPRAC has proposed a comment defining "reveal" to include exposing confidential information to technological systems where that creates a material risk. Read the terms of the specific plan before client facts go anywhere near it.
Does Astra for Law replace Westlaw or Lexis?
OpenAI does not position it that way. Its own announcement describes the legal search index as complementing the licensed content and specialist products firms rely on from providers such as Thomson Reuters. Treat it as an additional research surface with its own error profile, not as a substitute for a subscription your practice already depends on.
Sources
- OpenAI, Introducing Astra for Law, September 17, 2026
- The State Bar of California, Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law, issued November 16, 2023; revisions approved by the Board of Trustees May 14, 2026
- The State Bar of California, Proposed Amendments to the Rules of Professional Conduct Related to Artificial Intelligence, COPRAC approved March 13, 2026; comment period closed May 4, 2026
