Law, Finance, Health: What AI Labs' Industry Packages Add

On September 17, 2026 two frontier labs announced an industry package on the same day. OpenAI's Astra for Law adds a legal search index, written instructions and plugins to its flagship model. Anthropic's Life Sciences Verification Program adds nothing to the model at all: it is a verification process that removes safeguards for vetted biology teams. Both are sold as "for" an industry. They have almost nothing in common, which is the reason this census exists.

This page reads the vendor announcement for every named industry package from OpenAI, Anthropic and xAI, 15 in all, and takes each apart into six components: a curated index or dataset, written instructions or skills, tools and connectors, a safeguard or compliance change, an access gate, and a published price. It is for an executive in a regulated sector deciding whether a lab package, a specialist product or a self-built configuration fits. Google, Microsoft, Cohere and Mistral were checked and have no qualifying row; the reasons are in the methodology. This blog runs on Anthropic's models, and Anthropic's seven rows are decomposed by the same rule as the other eight.

A fourth count belongs in the second paragraph: the newest package in the census adds no capability. Anthropic's verification programme, which we covered in our post on what it changes, contains no index, no instructions and no tools; it is a gate that switches safeguards off for organisations that pass a review. OpenAI's legal package, which we took apart in our post on Astra for Law, is the opposite: everything except a gate that the page defines. Two labs used the same phrase in the same week for two different products, and a buyer who hears "industry version" should ask which of the six parts is in the box.

A row is a named package a lab announced for a vertical, built on its own model, with a dated primary. Partner products built on a lab's model are not rows. Two OpenAI rows are worth a note: "OpenAI for Government" dates from June 2025 but its current terms are the September 10, 2026 agreement with the US General Services Administration, and the Daybreak tiers are access levels to cyber-capable models announced in the same document, which we covered in our post on Astra's cyber threshold.

The first three components are what a package adds to the model's knowledge and behaviour. "Index" means data the lab curates and hosts. "Instructions" means written guidance, skills, plugins or templates. "Tools" means connectors into the systems the vertical already uses.

The second three components are about terms rather than capability: what the package changes in how the model is governed, who is allowed to buy it, and whether the vendor says what it costs.

Every qualifying row has at least one of five things: a retrieval index the lab builds and hosts; packaged instructions, skills, plugins or templates; connectors into the vertical's incumbent systems; a compliance setting the base product does not offer by default, such as HIPAA, FERPA or a data-training default; or an access decision about who may buy it or which safeguards apply once they have. That is what the phrase reliably means, and it is a useful set: most of it can be specified in a contract.

Three things it does not mean. It does not mean separate weights; no row ships them. It does not mean an independent benchmark against specialist products: the three rows that print a figure, OpenAI's legal research result, Anthropic's finance agent score and its life-sciences protocol score, each compare the package with another general model or with a human baseline, and none with Harvey, CoCounsel, Hebbia or any vertical incumbent. And it does not mean a price: 12 of 15 publish none, and no page read for this census mentions a service-level agreement, an indemnity or professional-liability terms.

Four questions the census makes answerable, each anchored to a row. Is this a different model? No; ask which of the six parts you are buying. Can I rebuild it? For a connector-and-plugin package, largely yes, and Anthropic ships its legal plugins and finance agent templates as open source; for a lab-hosted index, no, because the licensing and hosting are the product. How does it compare with the specialist I already pay? The vendors do not say. What will it cost? Assume you will have to ask, and note that the one published commercial-adjacent figure, on OpenAI's government page, is a $15 per user per month list price that the agreement waives.

The vertical guides on this site, such as our guide to agentic AI in banking and fintech and our analysis of the Claude legal plugin's market effect, cover the sector detail. If you want the six-part decomposition done for a package you are being sold, our AI transformation service does it before any contract is signed.