OpenAI shipped a legal search index covering more than 230 million13 URLs on September 17, 2026. Anthropic introduced Claude Docs and Claude Slides in beta on September 16, 202602. That is two frontier labs, 48 hours, and two products that compete directly with the companies paying them for API access.
Here is the uncomfortable part. Legal AI startups raised money on retrieval plus citation over case law. Office AI startups raised money on turning a chat into a document. Both of those pitches are now features inside the model vendor's own app.
One damaging admission up front: neither product has published adoption numbers. Astra for Law is gated behind a Trusted Access program for selected firms. Docs and Slides are in beta for all paid plans, rolling out first to Pro and Max subscribers. So this is a strategy read, not a scoreboard.
The plan for this piece is simple. Name the moat that just died. Walk through what each launch actually contains. Zoom out five years. Then give you a concrete rebuild path for a wrapper business this week.
Own the Consequence, Not the Corpus
Call it the Consequence Moat. The old moat was the corpus: you had the data, the index, the citations. The new moat is the consequence: you own what happens when the answer is wrong, and you own where it lands.
What two launches put on the board in 48 hours.
Think of it as a ledger with two columns. On the left, the things a lab can build with money and a crawler: an index, a chat surface, a document editor, an export to PowerPoint. On the right, the things a lab structurally cannot own: a firm's private matter history, malpractice liability, a live connection to the billing system, and the partner who signs the engagement letter.
OpenAI's own announcement puts the left column in plain view. Astra for Law "combines GPT-6 Astra, our latest and most powerful model, with settings, tools, and context tailored for professional legal work." That one sentence is the entire product of roughly a dozen funded startups.
The Consequence Moat says: if your value lives in the left column, you are a feature waiting to be absorbed. If it lives in the right column, the labs just became your best supplier. Sort your product honestly before you read another word.
Case Law Became a Feature Overnight
Start with the index, and how impermanent it turned out to be. According to RuntimeWire, Astra for Law's legal search index spans more than 230 million13 URLs across U.S. case law, statutes, regulations, court rules, and administrative decisions, with sources added daily. OpenAI describes it as one tool the model can call, not the product itself.
Read that again with beginner's mind. The thing legal AI vendors called their crown jewel is now a tool call. The lab also ships zero data retention on the API for eligible firms and excludes ChatGPT Enterprise usage from human review by default, according to OpenAI. Privacy, another pitch-deck slide, became table stakes in a single announcement.
Now the counterpositioning question: what did OpenAI deliberately not build? It named Harvey and Legora as API customers who "will be able to build on Astra for Law." It shipped 2611 ecosystem plugins connecting ChatGPT to Relativity and Clio. It said the index "complements the licensed content and specialist products firms rely on from providers such as Thomson Reuters."
That is a map of the right column, drawn by the lab itself. Relativity holds the discovery data. Clio holds the matters and the invoices. Thomson Reuters holds licensed editorial content and the annotations lawyers trust. OpenAI wants to sit next to those systems, not replace them. At least for now.
Anthropic ran the same play on a different floor of the building. Reuters reported that Anthropic combined chat and Cowork into a single interface and launched document and presentation tools, with Claude Design folded in as well. Docs export to Google Docs, Word, and PDF. Slides download as PowerPoint or PDF.
The pricing tells you the strategy. Usage of Docs and Slides is governed by your existing Claude plan limits, not by document count. Anthropic is not selling a docs product. It is raising the value of the seat it already sells, the way Costco reportedly sells a hot dog at a loss to keep you coming back and renewing the membership.
Reuters also placed this in a race: OpenAI launched ChatGPT Work in July 2026, combining its chatbot with Codex. Two labs, both preparing for potential IPOs by Reuters' account, are now fighting for enterprise seats with bundled surfaces. Seats are where recurring cash lives, and only cash is real. The rest is accounting.
Here is my read. The labs are not attacking their API customers out of malice. They are following gravity. The model is becoming a commodity input, so margin migrates to the surface where the work gets done. Nvidia nearly died in 1996 betting on a chip standard almost nobody wanted, then survived by owning the developer relationship with CUDA. The lesson was never "own the silicon." It was "own the layer people cannot leave."
Put it in pairs. Amateurs sell access. Leaders sell outcomes. A wrapper sells retrieval over public law. A durable vertical company sells the closed loop: intake, research, draft, partner review, filing, invoice, and the audit trail that survives a bar complaint.
The data is mixed on how far the labs will climb. OpenAI explicitly frames Astra for Law as "a new foundation for law firms and legal technology companies to build AI products." That could be sincere restraint. It could also be the standard platform posture right before the platform eats the ecosystem. Apple said something similar to third-party flashlight apps once.
It is unclear whether OpenAI will ever ship a full matter-management product, or whether Anthropic will build a real spreadsheet with formulas and version history. Building those means owning uptime for a law firm's billing run, and labs have shown little appetite for that kind of liability. My 70% confidence is that they stop at the surface and leave the plumbing to partners. I would not bet a company on the other 30% being zero.
Where the margin moved when the corpus stopped counting
The crown jewel became a tool call.
A legal index of more than 230 million URLs across case law, statutes, regulations and court rules now sits inside the model as one callable tool. Zero data retention for eligible firms arrived in the same announcement, turning privacy from a pitch-deck slide into table stakes.
Docs and Slides raise the seat, not the invoice.
Anthropic folded chat, Cowork and Design into one interface and priced document and presentation use against existing plan limits rather than document count. That is not a docs business. It is a defense of the subscription that was already sold.
Whether the labs climb higher is still open.
OpenAI named Harvey and Legora as builders and pointed at Thomson Reuters, Relativity and Clio through 2611 plugins. The author puts 70% confidence on the labs stopping at the surface, which leaves a 30% case nobody should bet a company against.
2031. Pull back five years and ask where the asymmetric risk sits. The downside for a retrieval wrapper is total: the lab bundles your feature into the subscription, and your revenue goes to zero in a quarter. The upside for a consequence-owning vertical company compounds: every matter processed makes the workflow data richer and the switching cost higher.
Consider three scenarios for 2031. In the first, the labs stay at the foundation layer, as OpenAI's language suggests, and vertical application companies grow because the base intelligence got cheaper and better every year. In the second, the labs climb into full suites, and the survivors are companies with regulated liability and system-of-record integrations too messy for a lab to want. In the third, the labs stall on enterprise trust, and integrators who own the deployment relationship capture most of the value.
Notice what all three share. Retrieval over public data is not the winning asset in any of them. The index is a flywheel input, and the labs now own the biggest flywheel: OpenAI adds sources daily to a corpus already past 230 million13 URLs. Competing on corpus size against that is competing on capex against a company raising for an IPO.
Salary buys furniture, equity buys your future. The analog here: API margin buys this quarter, workflow ownership buys the decade. Apply the AIM test to your own product. Actor: who is accountable when the output is wrong? Input: what data do you hold that the lab cannot crawl? Mission: what job is finished when your software runs, not just answered?
If the actor is the lab, the input is public, and the mission is "get an answer," you are in the left column. That position was fine in 2024. In September 2026 it became a countdown.
One more note. Maya, the illusion, took a specific form in this market: the belief that a clever retrieval pipeline was intellectual property. It was a head start, which is a different thing. Head starts are impermanent by definition. The companies that treat this week as a wake-up rather than a funeral will be the ones still operating in 2031.
Wire Your Wrapper Into Systems of Record
Show, don't tell. Here is what a founder running a domain wrapper can actually do before Friday, in order.
First, run the ledger. Open a blank document and list every feature in your product. Put each one in the left column (a lab could ship it with a crawler and a UI) or the right column (requires client data, client liability, or client systems). Be brutal. If a feature appears in OpenAI's Astra for Law announcement or Anthropic's Docs launch, it goes left.
Second, pick one system of record and build a real integration this week. Not a demo, a write-back. In legal, that means Clio or Relativity, the same tools OpenAI listed among its 2611 plugins. A system of record is the database a firm treats as the source of truth for matters, documents, or money. Reading from it is a feature; writing back to it with an audit trail is a moat.
Third, redesign your pricing around completed work. Stop charging per query or per seat if you can. Charge per matter closed, per contract reviewed and signed off, per filing submitted. Outcome pricing forces you to own the consequence, which is exactly the column the labs avoid.
Fourth, treat the labs as suppliers and use their new surfaces for your own operations. Claude Docs, Slides, and Design now live in beta inside the chat for paid plans, so draft your partner-facing deck in the same thread as your research and export it to PowerPoint. Anthropic's Claude for small business package ships 43 prebuilt workflows and 27 integrations; pick one workflow and one integration, then measure whether it saves an hour a week before adopting another. When you evaluate a model source like Meta AI's developer tools, treat it as an input, not a workflow. The workflow is your job now.
Fifth, expect breakage. Your first write-back to a practice management system will fail on a permissions error. The first outcome-priced contract will be mispriced. That is fine. Get your reps in, learn in public, and fix one thing at a time.
You do not need a CS degree for any of this. You need a spreadsheet with two columns, one integration that writes data, and the willingness to charge for finished work instead of answers. The labs just told you which column they are taking. Take the other one.
Sort your product into two columns, then go build the right one.
- Run the ledger. List every feature you ship and place each in the left column (a lab could build it with a crawler and a UI) or the right column (needs client data, client liability, or client systems). If a feature appears in OpenAI's Astra for Law announcement or Anthropic's Docs launch, it goes left.
- Ship one write-back integration. Pick a single system of record, in legal that means Clio or Relativity, the same tools inside OpenAI's 2611 plugins, and build a real two-way connection this week. A read-only demo does not create switching cost.
- Apply the AIM test in writing. Name the Actor accountable when output is wrong, the Input the lab cannot crawl, and the Mission that is finished rather than merely answered. If the actor is the lab and the input is public law, rewrite the roadmap before Friday.
Own the consequence, because the corpus is now rented.
In 48 hours, two labs converted the core pitch of a dozen funded startups into features inside their own subscriptions. Retrieval over public law does not win in any of the three 2031 scenarios, because the biggest flywheel now belongs to the company adding sources daily to a corpus already past 230 million13 URLs. What the labs structurally cannot own is private matter history, malpractice liability, and a live write path into the billing system. Sort your features honestly, pick one system of record, and build the closed loop that survives a bar complaint. The head start was never the intellectual property.