OpenAI Built Astra for Law on Free Public Caselaw. So Did We.
· LawDiver
OpenAI launched Astra for Law today — a GPT-6 configuration with a legal search index built on Free Law Project's CourtListener data, sold to Am Law 200 firms through an application form and a 30-page agreement. We built LawDiver on the same public sources. It is free during rollout, the cite checker hits ~97% accurate (100% on properly formatted cites; 100% on hallucinated caselaw), and the API and MCP server plug into any model — or any law app you are building.
OpenAI launched Astra for Law this morning — "a new foundation for law firms and legal tech companies to build AI products around their expertise."
It is a serious product, and the reporting on it is worth reading. It is also, underneath the announcement, a story about where legal data comes from. Because the single most important sentence in the launch is not about the model. It is this one: the caselaw in Astra's new legal search index comes from Free Law Project, the nonprofit that runs CourtListener.
That is the same public, unrestricted corpus LawDiver was built on.
So here is the honest comparison, written by a competitor, with our sources shown. Two teams reached the same conclusion about where American law lives. One of them is selling the result to the Am Law 200 through an application form. The other one — us — is giving it away during rollout, over an API and an MCP server that plug into whatever model you already pay for.
On this page
- What OpenAI actually shipped
- How they made it
- What 54% means
- How we made ours
- Same quarry, different building
- The two words that define this launch: contact sales
- Ours is free, and it is not a model
- The cite checker is the part we would build first
- Plug it into whatever you are building
- What Astra does that we do not
- Try it in five minutes
- Built for the rest of us
- Frequently asked questions
- Sources
- Related reading
What OpenAI actually shipped
Stripped of the launch language, Astra for Law is four things:
- A configuration, not a new model. It is GPT-6 Astra with legal-specific instructions for analysis and writing, custom response-length settings, and legal tooling attached. OpenAI says the configurations will expand over time. It shows up as GPT-6 Astra Law in ChatGPT and
gpt-6-astra-lawin the API. - A legal search index. More than 230 million URLs spanning U.S. case law, statutes, regulations, court rules, and administrative decisions, with sources added daily. The caselaw comes from Free Law Project and is said to cover more than 99.9% of published U.S. precedential case law.
- A distribution program called Trusted Access. Initially limited to selected U.S. firms — the Am Law 200 — with Zero Data Retention on the API, ChatGPT Enterprise usage excluded from human review by default, and delivery through ChatGPT and Codex. API access for builders is "coming soon," with Harvey and Legora named as early customers.
- An ecosystem. Twenty-six vendor plugins (Thomson Reuters, Harvey, Legora, iManage, Intapp, DeepJudge and others), nine community plugins, dozens of custom skills built by power users, plus ChatGPT for Word. Sullivan & Cromwell, Ropes & Gray, Cooley, Latham & Watkins, Wachtell, and Pillsbury are the named early builders.
There is no published price. To get it, you contact OpenAI.
How they made it
The engineering story is more interesting than the model story, and OpenAI has been straightforward about it: the model was not retrained for law. What changed is what the model can reach.
The index is the product. A frontier model asked to research American law without a legal index does what every frontier model does — recalls, searches the open web, and produces something Bluebook-shaped. OpenAI's fix was not a better brain. It was building a retrieval layer over primary law and pointing the model at it. That is the correct fix. It is the same fix we made, and the same fix Westlaw, Lexis, Harvey, and every other serious legal AI product has converged on.
The caselaw came from a nonprofit. Free Law Project spent years assembling and maintaining CourtListener, publishing quarterly bulk snapshots that Free Law Project states are free of known copyright restrictions. That is what made this launch possible on this timeline. The most valuable asset in OpenAI's legal launch is a public good maintained by a small nonprofit — and if you build in this space, you should know that, and you should support them.
"230 million URLs" is a URL count, not a case count. It covers caselaw, statutes, regulations, court rules, and administrative decisions across federal and state systems. It is a real coverage claim and a large one. It is not 230 million opinions, and nobody at OpenAI said it was. Read benchmark-adjacent numbers carefully in this category; the units move around.
The benchmark was outsourced. OpenAI reported results against the private validation set of Vals AI's Legal Research Bench — 200 U.S. legal research questions, rubrics authored and reviewed by practicing lawyers. Using an independent evaluator instead of a house benchmark is the right call, and more legal AI vendors should do it.
What 54% means
Here are the reported numbers, at the highest reasoning effort:
| Metric | Astra for Law | GPT-6 Astra + web search |
|---|---|---|
| Overall correctness (Vals Legal Research Bench, 200 questions) | 54.0% | 38.7% |
| Case references found | +24% | baseline |
| Relevant passages surfaced | up to +54% | baseline |
A 15.3-point jump from adding a retrieval layer is a large, real result. It is also the number the whole industry should sit with for a minute: the frontier legal configuration of the frontier model, with a purpose-built index over nearly all published U.S. precedential caselaw, is graded correct on 54% of research questions.
Three things follow.
First, this is a hard benchmark, not a broken product. Legal research questions with lawyer-authored rubrics are brutal to score. Comparisons across studies are not apples-to-apples — the Stanford RegLab work that measured Westlaw's AI-Assisted Research around 33% used a different benchmark and a different year — so treat 54% as "state of the art on this particular hard test," not "half of legal AI is wrong." But do not round it up to solved.
Second, answer length is doing some work. OpenAI reported that Astra's answers ran roughly twice as long as the baseline's. Longer answers surface more passages, and more passages score better on relevance rubrics. That is not cheating — depth is genuinely useful — but "more retrieved material" and "more correct analysis" are different achievements, and only one of them is what a partner signs.
Third, verification is the product. If the best available system is right 54% of the time on hard questions, then the thing standing between an AI research memo and a sanctions motion is not model quality. It is whether every citation in the output resolves to a real case that says what the brief says it says — checked mechanically, against a corpus, after the model is done talking.
That is the part we have spent most of our engineering on.
How we made ours
We published our full pipeline months ago, for exactly the reason this launch makes obvious: in a field where everyone claims "trusted content," provenance has to be showable. The short version:
Layer one — Harvard's Caselaw Access Project. Roughly 40 million pages of U.S. decisions digitized into about 6.7 million cases spanning around 360 years, available for open use. That is the historical backbone no live crawler could rebuild.
Layer two — CourtListener quarterly bulk data. The same Free Law Project source OpenAI just indexed: opinion clusters, opinions, dockets, courts, and citation maps, current through each quarter.
Layer three — reconciliation. Open does not mean clean. CAP and CourtListener disagree with each other, and both carry missing reporter citations, duplicate clusters that corrupt citation graphs, and gaps in recent regional reporter keys. We compared them row by row, updated texts, and resolved conflicts. This is the unglamorous half of the work and the half that decides whether you have a database or a dump.
Layer four — continuous harvest. A modified Juriscraper fleet across federal and all fifty state systems, plus OCR for the PDFs that arrive as scans, keeping the corpus current daily rather than quarterly.
Then the parts that are actually ours:
- EyeDiver — our citation extractor, which reads every opinion and finds every authority in the text, including the forms lawyers invent faster than parsers keep up with.
- The Bubbler — matches extracted cites to the authorities they point at, building the citation graph.
- Embeddings and a four-engine search stack — citation, case-name, lexical, and vector engines behind a router that produces a plan rather than a label, so
Miranda v. Arizona qualified immunity /s clearly establishedruns the name engine and the keyword engine at once instead of guessing which one you meant. Westlaw-style/s,/p, and/3connectors are verified in code against the candidate text, sentence by sentence — because a full-text index physically cannot express "same sentence," and faking it with a fixed word distance returns confidently wrong results. - Dual-LLM analysis and GoodLaw — our citator: treatment signals, negative history, and good-law status on every case, with no Shepard's content anywhere in it.
Scale today: on the order of 10 million opinions — majority, concurrence, and dissent each individually addressable — 78 million resolved opinion-to-opinion citation edges, 182 million in-text citations with character offsets and pin cites, 18 million reporter citations with normalized keys, a 133 GB Lucene index over the full text, twenty-five years of the U.S. Code, the CFR, and the Federal Rules stored the way lawyers cite them and versioned so a 2019 cite resolves to 2019 text, and about 525,000 lines of code — built by a very small team in months, which two years ago would have required a funding round.
Harvest is a job, not a quarterly ritual: crawlers walk more than 200 individual court websites around the clock, plus RSS and Atom feeds as an independent second signal, plus GovInfo, with quarterly bulk downloads as the historical backbone. Courts do not publish on a schedule — some in the morning, some after close of business, some in Friday batches — which is why yesterday's opinion is in the index today instead of in next quarter's dump.
And the number we care about most: our CiteCheck test bank, run across all 51 U.S. jurisdictions on 1,006 cases, lands at roughly 97% accurate when you combine fully accurate and mostly accurate — 100% on properly formatted cites, ~97% on slightly mangled ones the system can still identify and correct, and 100% on flagging hallucinated caselaw. No other product on the market meets these benchmarks. That is not a research-correctness score and it is not comparable to the Vals number; it measures something narrower and, for a lawyer about to file, more urgent: does this citation exist, is it this case, and is it still good law?
Same quarry, different building
| Astra for Law | LawDiver | |
|---|---|---|
| Caselaw source | Free Law Project / CourtListener | Harvard CAP + CourtListener + our own daily harvest |
| Coverage claim | 230M+ URLs; >99.9% of published U.S. precedential caselaw | ~10M opinions, 78M citation edges, 182M in-text cites |
| What was built | A retrieval index + model configuration | A corpus, a citation graph, a citator, a search stack, and products on top |
| Citator | Not announced | GoodLaw — treatment and negative history, no Shepard's content |
| Cite checking | Not announced as a discrete product | CiteCheck: ~97% accurate (100% clean cites; 100% hallucinated caselaw) on our 51-jurisdiction test bank |
| Model | GPT-6 Astra, configured | Whichever model you want; ours is a swappable ladder |
| Access | Trusted Access application, Am Law 200 first | Create an account |
| Price | Not published — contact sales | Free during rollout, rate card published unauthenticated |
| Agreement | Reported at 30 pages | Sign up |
Two teams looked at the same problem and reached the same first conclusion: the law itself is not the moat, because the law is public. What differs is everything after that — what you build on top, and who you will sell it to.
The two words that define this launch: contact sales
No price was published. Access is by application. The initial target is the Am Law 200. The access agreement, per OpenAI's legal products lead Jason Boehmig, runs about thirty pages — his line on the Zero Data Retention promise was that "it's definitely more complicated than the one sentence."
None of that is scandalous. It is how enterprise software has always been sold, and firms with ethical walls, client outside-counsel guidelines, and matter-level confidentiality obligations genuinely need those thirty pages.
But price the alternative for a second. Harvey — one of Astra's launch API partners — does not publish pricing either; reported buyer quotes put it in the range of roughly $1,200–$2,000+ per seat per month with seat minimums around 20–25, which pencils out to six figures a year before implementation. Westlaw and Lexis seats stack on top of that for most firms. Whatever Astra costs, it is being sold into that market, through that motion, at that scale.
Now count the lawyers that motion does not serve.
Solo practitioners. Two-person firms. Legal aid organizations. Public defenders. Small-firm litigators who file in state court every week and do their own cite checks at 11pm. In-house teams of one. Clinics. Law students. The developer with a legal AI idea and no procurement department to route a six-figure contract through. That is the overwhelming majority of American legal practice, and "contact sales, Am Law 200 first" is a polite way of saying not yet, and probably not at that price.
OpenAI's Boehmig said there will be more to share on access to justice in the future. We believe him, and we hope it is soon and it is real.
We are not waiting.
Ours is free, and it is not a model
Two things make LawDiver a different kind of answer to the same problem.
1. It is free during rollout
The API, the MCP server, and CiteCheck are live and free during the current rollout. Usage is metered against a published rate card and not charged. The rate card and every endpoint are visible without authentication, before you create an account or talk to anyone:
GET https://lawdiver.com/api/v1
That discovery document is unauthenticated on purpose. You should be able to see what an API offers and what it will cost before anyone asks you for a signature. Citation resolution (POST /api/v1/citations/resolve) is free permanently, so agent loops can resolve liberally before deciding what is worth a full check.
If the free rollout ends, the discovery document and the docs will say so before it does.
2. It is not tied to a model — it plugs into yours
This is the structural difference, and it matters more than price.
Astra for Law is a model configuration. Its legal index lives inside OpenAI's product. To use that index, you use that model, under that agreement.
LawDiver is the layer underneath. The corpus, the four search engines, the citation graph, the citator, and the cite checker are exposed as tools — over REST and over a hosted MCP server at https://lawdiver.com/mcp. Point any model at them:
- Claude, on desktop, web, or in Claude Code
- ChatGPT, via connectors
- Cursor and other MCP-aware editors
- Any client that implements the Model Context Protocol
- Your own agent, over plain REST, in any language
We are not ideologues about this. Our own AI research memos currently run an OpenAI model at the top of the ladder, through a router, with a fallback underneath — and that ladder is a configuration row we change after a bake-off, not an architecture. The model is a replaceable part. The corpus, the citation graph, and the citator are not.
That is the whole argument. If GPT-7 is better next quarter, you should be able to swap to it without re-buying American law. If your firm standardizes on Claude, your caselaw access should not care. If a court-adjacent client requires something on-prem, the retrieval layer should not be the thing that blocks you.
Rent the model. Do not rent the law.
The cite checker is the part we would build first
Everything above is about research. This section is about the thing that actually ends careers.
A legal AI product's reputation is not lost when it gives a mediocre answer. It is lost the first time it cites a case that does not exist. Since Mata v. Avianca the sanctions docket has filled with lawyers who trusted fluent output, and the pattern is always the same: the model did not invent a wild-sounding case. It produced a real reporter citation glued to the wrong caption — something that survives every eyeball test a busy associate can apply at 11pm.
Most "cite checkers" cannot catch that, because they ask the wrong question. They ask does this string look like a real citation? That question is unanswerable and the answer is worthless — a hallucinated cite looks exactly like a real one. That is the whole problem. The answerable question is:
Here is a citation string. Does it name a real case in a corpus we actually hold, what is its correct Bluebook form, and is it still good law?
Deterministic first, probabilistic as a net
Normalize 410 U.S. 113 to a canonical reporter key and identification is a single indexed lookup against 15.5 million citation-index rows and 18 million parallel-citation records — not a similarity score, not a model's opinion. The same document produces the same citation spans, the same reporter keys, and the same short-form resolutions on Monday that it produced on Friday. Reproducible, auditable, and safe to rebuild.
EyeDiver, our recognizer, started from Free Law Project's open-source eyecite and was substantially extended in-house — more citation types, tighter reporter resolution, short-form chains, and a reconciliation step against our own reporter-key index. It has been run across the entire corpus: roughly 10 million completed scans, 182 million in-text citations, 78 million resolved edges.
An LLM pass sits on top as a completeness net, never as the backbone, and its finds are stored as a distinct source so you can always see which citations came from the parser and which from the model. Determinism is the backbone; probability is the net. Neither philosophy alone survives contact with a real brief.
Verdicts that refuse to lie
This taxonomy is the product. Every citation comes back graded:
| Verdict | What it means |
|---|---|
valid | Exact reporter-key match and the case name, year, and court as written agree with that decision. Both halves required. |
name_mismatch | The reporter key resolved to a real case, but the caption, year, or court as written do not match it. This is the shape of a hallucinated citation — and it is never reported as valid. |
page_mismatch | The cite targets an internal pin-cite page rather than the decision's first page; correctedCitation supplies the proper form. |
likely_valid | Candidates found by name or partial signal, but the cite as written does not key cleanly — transposed volume, wrong page, misremembered reporter. Up to three ranked candidates, and we deliberately do not pick one. |
implausible | The reporter series or volume cannot exist. Strong fabrication signal. |
not_in_corpus | We searched a range we hold and found nothing. Actionable — but still not proof of fabrication. |
not_covered | We do not hold that range. Absence is evidence of nothing, and we say so. |
unverified | No source exists that can answer. Some statute sites return HTTP 200 for sections that do not exist, so a live probe there proves nothing. |
error | The check itself failed — emitted as a full row, never omitted, because an omitted row reads as "checked and fine." Not billed. |
Two design choices in there are worth the whole section. We never silently correct a citation to a case the author never read — that is the worst thing a cite checker can do, so ambiguous cites return ranked candidates and let a human choose. And confidence reaches exactly 1.0 only on a reporter-key match; there is no fuzzy 0.97 standing in for certainty.
The output is not "this looks fine." It is "this reporter cite is real, but it belongs to a different case; here is that case, here is its correct Bluebook form, and by the way it was questioned by the Ninth Circuit in 2021."
Whole briefs, not just citation strings
POST /api/v1/citecheck/document → { "jobId": "..." }
GET /api/v1/citecheck/jobs/:id/report
Upload a PDF or Word brief and every citation in it gets checked. Three things only show up on real filings:
- Short forms are bound, not re-resolved.
Celotex, 477 U.S. at 324is not an independent authority — it binds to a full cite earlier in the document and inherits that verdict. Resolving it standalone treats the pincite as a first page and reports "not found" for a correctly written Bluebook short form.Id.andsuprachains resolve the same way. This single failure mode makes most document checkers unusable on actual briefs. - Non-case citations are counted, not guessed at. Statutes, rules, and regulations are reported as skipped with a count rather than as "not found" from an index that never could have found them.
- Positions come back. Every result carries
docCharStartandanchorText, so you can highlight the exact span in your own viewer.
Measured: our CiteCheck test bank runs 1,006 cases across all 51 U.S. jurisdictions. Combining fully accurate and mostly accurate puts the headline number at roughly 97%. Split it the way a filing desk actually cares about:
- Properly formatted cites: 100% accuracy.
- Slightly mangled cites: ~97% accuracy — OCR noise, missing spaces, off-by-one pages, non-Bluebook punctuation. The system still identifies the intended authority and supplies the corrected form. The ~97% headline includes those mangled cites.
- Hallucinated caselaw: 100% accuracy on flagging. A real reporter cite glued to the wrong caption comes back
name_mismatch, nevervalid.
No other product on the market meets these benchmarks. Every jurisdiction's expected-versus-observed table is published in the repo, per case. Sunlight is the point.
The checker also knows the complete Bluebook for citation purposes. It can identify citations that do not strictly match Bluebook requirements and return the preferred Bluebook form — the same way it corrects slight mangling without pretending an ambiguous identity match is certain.
This is free during rollout, as a product at /citecheck and as an API endpoint. If you take one thing from this post: whatever model writes your memo — Astra, Claude, Gemini, ours — run the output through a corpus-backed checker before it goes anywhere near a filing. Fifty-four percent correctness is not a number you hand a judge.
Plug it into whatever you are building
Astra's index lives inside OpenAI's product. Ours is an ingredient. If you are building anything that touches American law, here is the part that matters: the API and MCP server expose the same stack our own products run on, and nothing in the terms tells you your product cannot compete with ours.
That last clause is not a throwaway. Incumbent data licenses routinely cap usage, forbid derivative indexes, or simply end with a "no" — vendors do not rush to arm their replacements. We can hand you this because none of it is licensed from a commercial research platform: no Westlaw feed, no Lexis feed, no scraped headnotes, no Shepard's or KeyCite extract. The inputs are public records and public government publications; the citation graph and the analysis are ours.
What you can wire it into
- A drafting or brief tool — cite-check the document on save, highlight
name_mismatchspans in the editor withdocCharStart, and show the correct Bluebook form inline. - A practice-management or DMS workflow — run
/citecheck/documenton every brief filed to the matter, store the report, and give the partner a clean audit trail. - An intake or triage bot — resolve the authorities a client mentions before a lawyer ever opens the file.
POST /api/v1/citations/resolveis free permanently, so resolve liberally. - A research agent —
/searchwith jurisdiction scope, read the good-law flags that ship on every hit,/cases/batchup to 50 at once,/cases/:id/cited-byto walk the graph forward. - A clinic, legal-aid, or access-to-justice tool — the population that "contact sales" does not serve, running on the same corpus the Am Law 200 is being sold.
- A courthouse-adjacent or internal app — plain REST, any language, no SDK required, no model dependency.
The loop that turns a demo into a product
- Constrain, then rank.
/searchwithsearchType: "auto", the right jurisdiction, and a date filter. The router runs name, keyword, and semantic together instead of guessing which one you meant. - Read the flags. Every hit already carries good-law status and a
matchExplanation— no second call to learn you are about to cite something overruled. - Pull what matters.
/cases/batchfor up to 50;/cases/:id/pdfwhen you need the court's own pagination. - Walk the graph.
/cases/:id/cited-byand/cases/:id/good-law. - Check your own output. Before the draft reaches a human, run
/citecheck/document. If anything comes backname_mismatch, your model just hallucinated and you caught it first.
Step 5 is the one we would build first, and it is the one almost nobody builds until after the incident.
The operational details builders actually ask about
One API key covers REST and MCP. Authorization: Bearer or X-API-Key, with a single gate that handles auth, rate limiting, and request-id assignment for every route — a new endpoint is structurally incapable of skipping it. Every response carries a requestId; every billable response carries a usage block showing what you were charged and why. Idempotency-Key is honored on billable POSTs: same key and same body replays the original answer, same key and a different body returns idempotency_conflict. Rate-limit headers on every authenticated call. Failed checks are not billed. Keys are scoped and revocable, so a connector can be switched off without touching anything else.
And the rate card is visible at GET https://lawdiver.com/api/v1 without authenticating — before you create an account, and before anyone asks you for a signature.
What Astra does that we do not
We would rather you trust this comparison than enjoy it, so here is the other side.
- Frontier reasoning is genuinely theirs. GPT-6 Astra at high reasoning effort is a better analytical engine than anything we ship in-house, and we do not pretend otherwise — we call frontier models for the analysis layer too.
- Breadth beyond caselaw. Their index spans statutes, regulations, court rules, and administrative decisions across the country. Our caselaw depth is our strength; our statute, regulation, and specialty-authority coverage is real but still expanding jurisdiction by jurisdiction.
- Firm workflow integration. iManage, Intapp, Relativity, Clio, HighQ, ChatGPT for Word, Codex. If your matter files, time entries, and DMS are the bottleneck, that ecosystem is a real answer and we do not have one.
- Enterprise contracting. Zero Data Retention on the API, human-review exclusions, and a negotiated agreement built with Latham & Watkins on ethical walls and client instructions. That is exactly what a large firm's GC and CISO need, and it is thirty pages for a reason.
- Independent benchmarking. They ran Vals. That is a higher standard of proof than most of this industry offers, including, so far, us — our published numbers are our own test bank, run by us.
If you are a large firm with a procurement team, an innovation budget, and a DMS integration roadmap, Astra for Law may well be the right purchase. Go look at it seriously.
If you are everyone else, keep reading.
Try it in five minutes
MCP — put verified caselaw inside the assistant you already use:
{
"mcpServers": {
"lawdiver": {
"url": "https://lawdiver.com/mcp",
"headers": {
"Authorization": "Bearer YOUR_LAWDIVER_API_KEY"
}
}
}
}
Your assistant gets case search, opinion retrieval with full text and structured metadata, citation verification with treatment signals, good-law status, and per-session jurisdiction scoping — so it cannot wander into authority that does not bind your court.
REST — the same stack, for your own agent:
POST https://lawdiver.com/api/v1/search
{
"query": "\"clearly established\" /s qualified immunity NOT prison",
"jurisdiction": { "type": "federal_circuit", "circuit": "11" },
"searchType": "keyword",
"limit": 20,
"filters": { "dateFrom": "2015-01-01", "publishedOnly": true },
"include": { "caseCard": true, "opinionText": true, "goodLawReport": true }
}
Cite-check — the call we would make first:
POST https://lawdiver.com/api/v1/citecheck/cite
{ "citations": ["Smith v. Jones, 570 U.S. 744 (2013)", "123 F.4th 999 (5th Cir. 2024)"] }
Anything that comes back name_mismatch is a real reporter cite glued to the wrong caption — the exact signature of a hallucinated citation.
Every response carries a requestId. Every billable response carries a usage block showing what you were charged and why — including, right now, zero. One API key covers both MCP and REST.
No sales call. No 30-page agreement. No Am Law 200 requirement. Create a free account and you have a key in about a minute.
Built for the rest of us
Today's launch settles an argument we have been having with this industry since we started: American primary law is public, and a product built on it does not require a Westlaw license or a nine-figure model lab. OpenAI just validated that thesis at the largest possible scale, using the same nonprofit bulk data we did.
The part they left open is who gets to use the result.
Astra for Law is an excellent product aimed at the top 200 firms in the country and the vendors who sell to them. It will be very good, and it will be priced accordingly, and both of those things are fine.
LawDiver is aimed at the other 99%. Ten million opinions, a real citation graph, a citator, a cite checker that has been run against 1,006 cases in 51 jurisdictions, an API and an MCP server that work with any model you like — free during rollout, with the price list published before you sign anything.
LawDiver. Built for the rest of us.
Frequently asked questions
Is Astra for Law a new model? No. It is OpenAI's GPT-6 Astra configured for legal work — legal-specific instructions, response settings, and a dedicated legal search index. It appears as gpt-6-astra-law in the API.
Where does Astra for Law's caselaw come from? From Free Law Project, the nonprofit behind CourtListener, as part of its legal search index of 230M+ URLs. That is the same public, unrestricted bulk source LawDiver used as one of its layers.
How much does Astra for Law cost? OpenAI has not published pricing, eligibility, or usage limits. Access is through the Trusted Access program, initially for selected U.S. firms, by contacting OpenAI. For scale, launch partner Harvey is reported at roughly $1,200–$2,000+ per seat per month with seat minimums.
How much does LawDiver cost? The API, MCP server, and CiteCheck are free during the current rollout. Usage is metered against a rate card that is published unauthenticated at https://lawdiver.com/api/v1, and citation resolution is free permanently.
Can I use LawDiver with ChatGPT? Yes — via connectors to our hosted MCP server, the same way you would connect it to Claude, Claude Code, Cursor, or any MCP client. The REST API works from any language and any agent framework.
Does LawDiver use OpenAI models? Yes, among others, on a swappable provider ladder. That is the point: the model is a replaceable part. The corpus, citation graph, and citator are what you should not have to re-rent when you change models.
Can I build my own law app on the LawDiver API? Yes. Search, case retrieval, the citation graph, good-law status, and cite checking are all exposed over REST and MCP, in any language, with no SDK required. None of the corpus is licensed from a commercial research platform, so there is no clause telling you your product cannot compete with ours.
How is your cite checker different from asking an LLM to check citations? An LLM checking citations is the same failure mode that produced the bad cite. Ours resolves each citation against a real corpus by normalized reporter key, then grades it — including name_mismatch, a real reporter cite attached to the wrong caption, which is what a hallucinated citation actually looks like. It knows the complete Bluebook, flags and corrects cites that do not strictly match Bluebook form, and binds id., supra, and short forms to their antecedents instead of reporting correctly written Bluebook short forms as "not found." On our 51-jurisdiction test bank that works out to ~97% accurate overall (100% on properly formatted cites; 100% on hallucinated caselaw).
What does 54% correctness mean for me? That even the best legal AI available today is wrong often enough on hard research questions that mechanical verification is not optional. Whatever model writes your memo, cite-check the output against a real corpus before it goes anywhere near a filing.
Sources
- Introducing Astra for Law — OpenAI
- OpenAI Releases Astra for Law, A GPT-6 Model Tailored for Legal Work, Targeting Large Firms and Tech Vendors — LawSites
- Breaking news: OpenAI unveils Astra for Law — Legal IT Insider
- OpenAI Introduces Astra for Law With Legal Search and Trusted Access — Unite.AI
- OpenAI introduces new AI model for law firms — ABA Journal
- Vals Legal AI Report — Vals AI
- Vals AI's Latest Benchmark Finds Legal and General AI Now Outperform Lawyers in Legal Research Accuracy — LawSites
- CourtListener bulk data — Free Law Project
- Caselaw Access Project — Harvard Library Innovation Lab
- Harvey AI Pricing in 2026: What It Actually Costs — The Legal Prompts