OpenAI shipped a legal vertical and the plugin partner list is the legal AI market — including the two vendors whose whole product was 'GPT, for law'

What shipped

In June 2026, OpenAI hired Jason Boehmig — co-founder and former CEO of the contract lifecycle company Ironclad, a former Fenwick & West corporate attorney — to build products for the legal sector. Three and a half months later, the vertical exists.

Astra for Law is GPT-6 Astra configured as a legal product. Three layers sit on top of the base model:

Domain configuration. Legal-specific instructions covering how to write and analyse — treatment of authority, surfacing unsupported assumptions, citation behaviour. Sullivan & Cromwell partner John Savva, quoted in the launch coverage, singles out the model’s “sensitivity to authority,” which is the phrase that matters in a field where a plausible-sounding citation to a vacated opinion is a malpractice event rather than a bad answer.

A legal search index. More than 230 million URLs of US case law, statutes, regulations, court rules and administrative decisions, with new material added daily. LawSites identifies CourtListener as a data source.

Firm governance. Information permissions and ethical walls, which Latham & Watkins is named as having helped design. Michael Rubin of Latham framed it as building on the firm’s “broader investments in AI development and governance.”

Four named firms brought concrete builds to launch: Sullivan & Cromwell (an agreement analyser), Ropes & Gray (an M&A diligence system), Cooley (a capital markets tool for its GO Public programme), and Latham on the permissions architecture. Pillsbury is also named among early access firms.

That is a real product with real firms behind it. The interesting part is what OpenAI chose to own.

The index is the product

Read the benchmark carefully, because it is doing something unusual for a launch number.

Astra for Law passed the evaluation’s overall correctness check on 54.0% of questions, compared with 38.7% for GPT-6 Astra using web search alone — a 40% relative improvement.

Both figures come from the same model. This is not Astra against a weaker competitor, or Astra against last year’s model. It is Astra with OpenAI’s legal index against Astra with a general web search tool, across 200 legal research questions, with supporting numbers of 24% more reference cases surfaced and 54% more relevant passages pulled from the correct opinions.

OpenAI is, in effect, publishing the size of the gap that retrieval alone accounts for in legal work — and then keeping the thing that closes it.

For the last two years the pitch of the entire legal AI category has been some version of we have taken a frontier model and made it safe and useful for law. The taking-and-making was mostly retrieval engineering: a clean corpus, citation resolution, authority checking, treatment signals. That was the labour. That was the moat.

It is now a configuration of the platform.

The plugin list is the market

OpenAI’s framing is cooperative, and Boehmig said so directly: the intent was to lead with an ecosystem-forward approach. The evidence is real — 26 partner-built plugins at launch, from Thomson Reuters, Intapp, Harvey, Legora, DeepJudge, iManage, Relativity and Clio, plus nine community plugins from LegalQuants, LECG and Skills.law, and 47 custom skills contributed by legal power users.

Then look at where Harvey and Legora appear.

They are plugins inside ChatGPT. They are also API customers who will build Astra for Law into their own products. Two weeks earlier, both had been publicly enthusiastic about the underlying model — Harvey’s head of applied research called Astra “a significant quality improvement over GPT-5.6 Sol across complex legal tasks,” and Legora reported a financial-statement tie-out across 41 documents in a single run that caught all four errors it had planted, including a £500,000 gap hidden in the revenue note.

Those are genuine capability wins for both companies. They are also the position of a firm whose product quality is set by a supplier that has now entered its market, with the supplier’s own retrieval layer attached.

This is not a prediction that Harvey dies. Harvey retains things Astra for Law does not attempt — Vault’s review of up to 100,000 documents at once, codified firm workflows, specialist agents for immigration, tax and M&A, and the unglamorous work of deploying software inside a partnership. At roughly $1,000–$1,200 per lawyer per month with 20-seat minimums, though, a meaningful share of that price has been buying research quality. Research quality is the line item that just moved downstream.

Thomson Reuters’ response is the most precise sentence any vendor produced this week. CTO Joel Hron supplied a partner quote about legal professionals needing “trusted intelligence, relevant enterprise context, and governance required for high stakes work” — while making clear that CoCounsel remains the trusted professional AI system. That is a vendor drawing a boundary around the part OpenAI’s public-record index does not reach: Westlaw’s licensed corpus, headnotes, KeyCite treatment. It is a correct boundary. It is also a smaller boundary than the one Thomson Reuters was defending a year ago.

The pattern this confirms

This desk has now watched the same manoeuvre in three verticals in three weeks.

In healthcare, OpenAI connected ChatGPT to Epic’s EHR with nine official health data sources and five named health systems — where the finding was that frontier labs have stopped trying to be your system of record and started competing to be the reading layer over everyone else’s. In enterprise software, Salesforce moved the reasoning into a partner model and left the meter blank. Now in law, the same shape, with one upgrade: OpenAI did not connect to someone else’s legal database. It built the index.

The buyer-facing consequence is a change in the question that survives contact with next quarter.

“Which model does your product use?” has a short half-life — the answer changes when the vendor’s supplier ships, and it is rarely the vendor’s own achievement. The question with a longer half-life is: what does this vendor own that the model provider cannot configure into existence? Proprietary or licensed content. Write-access integrations into systems of record. Regulated workflow and the accountability attached to it. Deployment depth inside an organisation that does not want to change how it works.

If the answer is a good system prompt and a retrieval pipeline over public sources, that is now a feature of the platform. In law, as of 17 September, it literally is.

There is a second-order caution attached. A plugin directory is a dependency, and OpenAI has retired platform layers before — the Assistants API sunset closed exactly twelve months after notice, with five further retirements scheduled before 20 January 2027. Building a legal workflow that assumes a ChatGPT plugin surface will exist in 2029 is a bet on OpenAI’s roadmap, not on your own.

What is actually gated, and what is not

The capability is narrower than the announcement’s reach.

Availability is invitation-only. Trusted Access, selected firms, delivered through ChatGPT and Codex. Named early adopters are Am Law 200. API access is “coming soon” with no date.

No price exists publicly — not for Trusted Access, not for the eventual API. Given that GPT-6 Astra itself lists at $10 in / $50 out per million tokens, and that legal research is a long-context, high-retrieval workload, the API price is the single number that determines whether this compresses the category or just adds a premium tier at the top of it.

The data terms are better than the default. Legal IT Insider reports the Trusted Access arrangement includes Zero Data Retention, with ChatGPT Enterprise usage excluded from human review, under an agreement running to roughly thirty pages. ZDR matters here more than in most verticals — privileged material cannot sit in a retention window — and it is worth noting that OpenAI has been building toward exactly this, having previewed misuse detection it says stays ZDR-compatible in August. Firms outside the US should also confirm where this runs, given the regional gaps that shipped with Astra itself.

The corpus is US-only. Case law, statutes, regulations, court rules and administrative decisions of the United States. For any practice whose research burden is English, EU or commonwealth law, the index that produced the 54.0% figure does not apply.

What to do about it

For large firms: if you are inside Trusted Access, the governance work Latham did is the part worth copying — ethical walls and information permissions are the difference between a research tool and a conflicts incident. If you are not, this is not yet procurable, and the correct move is to hold.

For firms currently negotiating or renewing legal AI contracts: the leverage changed on 17 September and your vendor knows it. Ask explicitly which of the four defensible surfaces — licensed content, workflow depth, DMS integration, governance — the price is buying, and decline to pay a research-quality premium that a platform configuration now substitutes for. Avoid long lock-ins priced chiefly on research quality until the Astra for Law API price is published.

For small firms and solos: nothing reaches you directly. The realistic upside is that the cost floor for building a credible legal research tool falls, and the tools that do serve you improve on someone else’s capital. Watch for the API price; that is the trigger.

For everyone buying any AI tool, in any category — the productivity stack included: run the ownership test. Model choice is not a moat, and increasingly it is not even a decision. Ask what is left above the model, and make the vendor answer in nouns.

Sources: OpenAI — Introducing Astra for Law; LawSites; Legal IT Insider; Artificial Lawyer; Artificial Lawyer — Harvey + Legora on GPT-6 Astra; LawSites — Boehmig joins OpenAI. Benchmark figures and the 230M+ URL index size are OpenAI’s own published claims, reported consistently across the four trade outlets above; they have not been independently reproduced. No pricing has been disclosed.

Questions or corrections? Email Pick Right. Want the full list? See all news.