Build Independently confirmed 3 publishers 3 min read Published Sep 17, 2026, 4:37 PM PDT Updated Sep 18, 2026, 11:11 PM PDT

OpenAI bundles a 230 million-URL legal index behind the model ID gpt-6-astra-law

The legal toolkit ships as its own entry in the ChatGPT model picker and its own API string. Firms weighing it against retrieval stacks they already built still have no price to compare.

The Engineer · Build desk

Send a correction

Illustration accompanying OpenAI bundles a 230 million-URL legal index behind the model ID gpt-6-astra-law — Generated illustration

What happened

OpenAI launched Astra for Law on September 17, packaging its GPT-6 Astra flagship with legal instructions, longer-work settings and a Legal Search Index for law firms and legal software vendors.
Access begins with selected firms through a Trusted Access program inside ChatGPT and Codex, and OpenAI says API access will follow.
OpenAI shipped 26 partner-built plugins and 47 community plugins alongside firm-built systems from Sullivan & Cromwell, Ropes & Gray and Cooley.

Why it matters

Who has to decide decision A team that already assembled legal retrieval now picks between a corpus it controls and a configuration OpenAI maintains and can change underneath it.
Who's exposed exposure Harvey, Legora, Thomson Reuters and iManage gain ChatGPT distribution while OpenAI supplies the model and the maintained legal configuration beneath their products; runtimewire argues standardization pushes their differentiation into content, access controls and audit trails.
What doesn't add up contradiction Dev.to reports the index as making authoritative material more accessible inside an AI workflow, with lawyer review still required; runtimewire says the URL count does not establish how duplicates, subsequent history or conflicting authorities are handled. Those mechanics decide whether a cite can be relied on.

Next story · Product · Oct 1

OpenAI holds GPT-6.1 Astra back after the model kept working past what users asked

Confirmed4 publishers

The detail that touches an engineering plan is the model string. OpenAI lists the legal version as GPT-6 Astra Law in the picker and as gpt-6-astra-law in the API ^[21]. A distinct ID means another model to qualify. Every eval, regression set and citation check that currently guards your pipeline gets rerun.

What the bundle covers is the cheap half of a legal RAG stack. In its launch thread on X, OpenAI described instructions for legal analysis and writing, settings intended for longer and more thorough work, and a new Legal Search Index ^[9]. Instructions and settings are the parts a competent team already keeps in version control. The index is the part that costs real money to build and more to keep current. It spans US case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs ^[5]. It draws on CourtListener data that OpenAI says covers a substantial share of published US precedential law ^[6].

The 230 million figure counts addresses in a crawl. Runtimewire's reading is that it describes the size of the search surface and does not establish how the index handles duplicate documents, subsequent legal history, conflicting authorities, or the difference between controlling and persuasive precedent ^[13]. For the number to transfer to your matters, the corpus would have to cover your jurisdictions and be current to the last filing date you care about. It would also have to collapse the same opinion appearing at forty addresses into one holding.

GPT-6 Astra itself shipped on September 3rd as OpenAI's model for demanding professional and computer-based work ^[24], fourteen days before the legal packaging ^[27]. Its API documentation lists a 1.05 million-token context window, tool support including web and file search, and reasoning settings from low to max ^[25]. Given the description of settings for longer and more thorough work ^[9], I would expect the legal preset to sit near the top of that range, though that is inference from the general docs rather than a published value.

The buy side cannot be priced yet. OpenAI has not published pricing, a broad API launch date or the eligibility requirements for Trusted Access ^[23]. The Zero Data Retention commitment is stated for API access by eligible law firms ^[10]. The first deployments run in ChatGPT and Codex ^[3]. If your index is CourtListener-derived and your product differentiates on workflow, I think the migration is worth planning for once API access and retention terms are contractual. If what you sell is subsequent-history logic and an audit trail you can defend to a partner, this bundle asks you to swap an inspectable component for an opaque one.

For vendors the position is more awkward. OpenAI shipped 26 partner-built plugins and 47 community plugins for legal work in ChatGPT ^[7], 73 in total ^[22]. Thomson Reuters, Harvey, Legora and iManage are among the named partners ^[8], and OpenAI says API customers such as Harvey and Legora will be able to build on Astra for Law ^[16]. Runtimewire argues that if the model layer becomes standardized, legal technology vendors will have to distinguish themselves through proprietary content, access controls, audit trails and workflow design instead of basic model access ^[28]. The three firm-built systems OpenAI named point at the same class of work. They are Sullivan & Cromwell's agreement analyzer, built with OpenAI engineers, Ropes & Gray's M&A diligence system, and Cooley's GO Public workflow for preparing companies to go public ^[12].

What to watch

Published API pricing for gpt-6-astra-law, and whether Zero Data Retention becomes a contractual term or stays a stated commitment.
Documentation of the index itself: update cadence, deduplication, subsequent-history treatment and jurisdiction coverage.
Whether Harvey and Legora build on Astra for Law's retrieval or keep their own index underneath their products.

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence55
Adoption35
Hype gap+25
Incentives70
Confidence62

Perspective Coverage

3 publishers

Builder: Builder 38%
Operator: Operator 30%
Investor: Investor 32%

Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

[1]
OpenAI announced Astra for Law, a legal AI foundation combining its GPT-6 Astra flagship model with a legal-specific toolkit for law firms and legal technology companies.
Reported SupportedView cited source
[2]
OpenAI launched Astra for Law on September 17th.
Reported SupportedView cited source
[3]
The initial rollout is limited to selected firms through OpenAI's Trusted Access Program in ChatGPT and Codex, while API access is described as coming soon.
Reported SupportedView cited source
[4]
OpenAI says API access will follow, without giving a launch date.
Reported SupportedView cited source
[5]
OpenAI says the Legal Search Index searches U.S. case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs.
Reported SupportedView cited source
[6]
OpenAI says the index leverages CourtListener data to cover a substantial share of published US precedential law.
Reported SupportedView cited source
[7]
OpenAI launched 26 partner-built plugins and 47 community plugins for legal work in ChatGPT.
Reported SupportedView cited source
[8]
The named plugin partners include Thomson Reuters, Harvey, Legora and iManage.
Reported SupportedView cited source
[9]
OpenAI said in a five-post thread on X that the offering combines GPT-6 Astra with instructions for legal analysis and writing, settings intended for longer and more thorough work, and a new Legal Search Index.
Reported SupportedView cited source
[10]
The announced package includes Zero Data Retention on API access for eligible law firms.
Reported SupportedView cited source
[11]
The 26 partner-built plugins include Relativity, Clio and HighQ/CoCounsel Legal through Thomson Reuters.
Reported SupportedView cited source
[12]
OpenAI named three systems developed with large U.S. law firms: Sullivan & Cromwell built an agreement analyzer with OpenAI engineers, Ropes & Gray developed an M&A diligence system, and Cooley built GO Public, a workflow for preparing companies for initial public offerings.
Reported SupportedView cited source
[13]
According to runtimewire, the figure of 230 million URLs describes the size of the search surface and does not establish how the index handles duplicate documents, subsequent legal history, conflicting authorities or differences between controlling and persuasive precedent.
Reported SupportedView cited source
[14]
The launch packages a frontier model with legal retrieval, domain instructions and controls instead of leaving firms and software providers to assemble those elements independently.
Reported SupportedView cited source
[15]
Astra for Law offers firms a maintained legal configuration instead of leaving each customer or software vendor to assemble prompts, retrieval systems and model settings around the underlying model.
Reported SupportedView cited source
[16]
OpenAI says API customers such as Harvey and Legora will be able to build on Astra for Law.
Reported SupportedView cited source
[17]
The announced toolkit includes a legal search index, specialized instructions for legal analysis and writing, legal-grade trust and controls according to OpenAI, a Trusted Access Program for eligible law firms, Zero Data Retention on API access, and plugins and integrations connecting ChatGPT with established legal software.
Reported SupportedView cited source
[18]
OpenAI described Astra for Law as supporting lawyers' expertise and judgment, stopping short of presenting the system as a substitute for either.
Reported SupportedView cited source
[19]
OpenAI cites work with firms and vendors including Sullivan & Cromwell, Harvey, Thomson Reuters and Latham & Watkins.
Reported SupportedView cited source
[20]
Dev.to reports that research quality still depends on the underlying sources, the matter at hand and lawyer review, and that the index is intended to make authoritative legal material more accessible within an AI-assisted workflow.
Reported SupportedView cited source
[21]
The offering appears as GPT-6 Astra Law in the model picker and as gpt-6-astra-law in the API.
Reported SupportedView cited source
[22]
Astra for Law launched with 73 legal plugins in ChatGPT in total.
Derived Supported
[23]
The announcement does not provide public pricing, a broad API launch date or detailed eligibility requirements for Trusted Access.
Reported SupportedView cited source
[24]
GPT-6 Astra launched on September 3rd as OpenAI's model for demanding professional and computer-based work.
Reported SupportedView cited source
[25]
OpenAI's API documentation for GPT-6 Astra lists a 1.05 million-token context window, tool support including web and file search, and reasoning settings ranging from low to max.
Reported SupportedView cited source
[26]
Runtimewire says OpenAI is giving legal vendors a distribution channel inside ChatGPT while building its own legal search, model configuration and workflow layer alongside them.
Reported SupportedView cited source
[27]
Fourteen days separated the GPT-6 Astra launch and the Astra for Law launch.
Derived Supported
[28]
Runtimewire says that if the model layer becomes standardized, legal technology vendors will have to distinguish themselves through proprietary content, access controls, audit trails and workflow design rather than basic model access.
Reported InsufficientView cited source

Sources

3 independent publishers whose own reporting we read for this story.

dev.to
2 articles · September 18, 2026
OpenAI Astra for Law Brings GPT-6 Astra to Legal Research and Workflow Building
runtimewire.com
1 article · September 17, 2026
OpenAI launches Astra for Law with a 230 million-URL search index
the-decoder.com
2 articles · September 18, 2026
OpenAI takes aim at the legal market with Astra for Law

Clarity analysis

How the coverage leans and what the record actually supports. Read the methodology

Perspective Coverage

How much of the coverage speaks to builders, operators and investors. It measures emphasis, not whether the story is right.

Builder38%

OpenAI published the model ID before the endpoint. gpt-6-astra-law is in the announcement, but the only access is ChatGPT and Codex for selected Trusted Access firms, and dev.to and runtimewire both note the API date is open. For now the buildable surface is the 26 partner plugins and 47 community plugins, plus whatever Harvey and Legora ship once the endpoint opens. Runtimewire asks the question a retrieval engineer would ask first: how the index handles duplicate documents, later history and conflicting authority. Size is the only property of the index OpenAI quantified.

Operator30%

A firm comparing this against the retrieval stack it already built cannot find a price anywhere in the record. Dev.to is the source that says so plainly, and adds that Trusted Access eligibility is undefined. Zero Data Retention is offered to eligible firms on API access, which is the part that has not shipped. OpenAI still puts verification on the lawyer: it describes the system as supporting a lawyer's expertise and judgment, and both write-ups say authorities, currency and traceability still need checking by hand.

Investor32%

Harvey, Legora, Thomson Reuters and iManage are in the plugin list and are also the companies whose layer OpenAI has now built. Runtimewire draws that out: partners supply content, integrations and finished applications, OpenAI supplies the model, interface and maintained legal configuration. The-decoder carries the only score, OpenAI's own Vals AI Legal Research Bench run at 54 percent against 38.7 percent for the base model with web search, and notes Anthropic is pushing into the same market. A separate dev.to post on the same launch week puts Polymarket's September best-model odds at 98 percent Anthropic and 1 percent OpenAI.

Reality

Evidence, adoption, hype gap, incentives and confidence, each read independently from the current record on the scale shown beneath it. How these are measured.

Evidence55: 0100
One announcement, three retellings, one self-run test
The 230 million URLs, the Trusted Access gating, the six toolkit items and the model ID all come from OpenAI's five-post thread on X and its announcement page. Runtimewire and dev.to restate that material carefully and disagree with none of it. The-decoder supplies the only measured figure, and says OpenAI ran the test: 54 percent of 200 Vals AI Legal Research Bench questions against 38.7 percent for the base model with web search. OpenAI is, so far, the only party with a published result.
Adoption35: 0100
Named pilots, gated access
Three real workflows exist inside Sullivan & Cromwell, Ropes & Gray and Cooley, built with OpenAI engineers, and 73 legal plugins were live in ChatGPT on day one. Both numbers measure supplier and design-partner activity. Everything else is gated to selected firms in ChatGPT and Codex. OpenAI has not opened the API or disclosed user or matter counts from any firm.
Hype gap+25: −100 understated0 aligned+100 overstated
Substrate language, pilot substance
Runtimewire's reading of OpenAI's bet is that a common foundation becomes the default substrate for legal workflows. What shipped is a gated configuration with an unpriced roadmap. The retrieval claim carries most of the gap: 230 million URLs sizes the corpus and tells a lawyer nothing about duplicates, subsequent history or which authority controls, and runtimewire is alone in pressing it. The counterweight sits in dev.to's other post that week, which prices OpenAI at 1 percent on Polymarket's September best-model question while launch coverage ran wall to wall.
Incentives70: 0100
Everyone in the list gains from being in the list
OpenAI is the source and the seller. The partners it names, Harvey, Legora, Thomson Reuters, iManage, Relativity and Clio, gain ChatGPT distribution from appearing, and runtimewire is explicit that OpenAI is building underneath the products those same vendors sell. On the publishing side, the dev.to piece ends with a pitch for Scalevise's AI consultancy and a request-a-consultation link, and the-decoder's page ends in a subscription offer.
Confidence62: 0100
Solid on what shipped, thin on what it does
That the product exists, what it is called, who helped build it and how narrowly it is available are all well attested and consistent across runtimewire, dev.to and the-decoder. Everything a buyer would weigh, price, eligibility, retrieval behaviour under conflicting authority, sits on OpenAI's word or on nothing. I would expect the first genuine test to come from a firm outside the design-partner group.

Topics and entities

Topics

Vertical Model Packaging
Data retention controls
AI Platform Strategy
Legal research retrieval
Legal AI