On 30 June 2026, Anthropic launched Claude Science. Read the description carefully, because the important word is what it isn’t: it isn’t a new model. It’s a finished product — an AI workbench for scientists that runs on existing Claude models and bundles literature search, code, figures, compute jobs, and reproducible artifacts into one environment, with 60-plus curated skills across genomics, proteomics, structural biology and cheminformatics, and reviewer agents that check citations and calculations. MIT Technology Review called it Anthropic’s newest flagship product. On the same day, the company announced it would run its own drug-discovery program aimed at neglected diseases.

For two years the frontier labs sold one thing: raw intelligence, metered by the token, for everyone else to build products on top of. Claude Science is a lab deciding to build the product itself. That’s not a feature release. It’s a move up the stack — and if you make your living somewhere above the token layer, it’s worth understanding exactly which floor the elevator just stopped on.

For the rest of us: what “the stack” means

Think of the AI industry as a building with three floors.

The ground floor is raw intelligence — the model, sold through an API and billed per token. Powerful, but unfinished; on its own it’s an engine, not a car.

The middle floor is platforms and tools — the scaffolding that turns the engine into something usable: coding assistants, agent frameworks, developer tooling.

The top floor is finished vertical products — software built for one industry’s actual job: a workbench a scientist opens and uses, a tool a lawyer or radiologist relies on. This floor is where most of the industry’s startups and enterprise projects live, because it’s where raw intelligence turns into something a specific customer will pay for.

For two years the labs occupied only the ground floor and let everyone else build upstairs. Claude Science is a lab taking the elevator to the top floor and opening a shop there.

The lab’s rising footprint

The pattern isn’t unique to science, which is what makes it strategic rather than anecdotal. Anthropic already climbed to the middle floor with developer tooling; Claude Science is the top floor. Each release, the lab’s footprint expands upward into space that used to belong to its own customers. The company that sells you the engine is now selling finished cars in your showroom — using an engine it can tune better than you can, because it built it.

Who gets squeezed

Two groups feel this directly, and it’s worth being precise about which.

The first is thin-wrapper startups — companies whose product is mostly a nice interface and some prompt engineering around a frontier model. When the lab ships the same capability as a first-party feature, the wrapper’s core value proposition becomes a checkbox on its supplier’s product page. “Your product is now a feature of the company you depend on” is the oldest platform risk there is; it just arrived in AI.

The second, and less obvious, is enterprises midway through building their own vertical AI. If you’re a pharma company or a bank with a two-year internal program to build a bespoke AI workbench, the launch of a credible first-party product from your model vendor is a genuine build-versus-buy shock. The thing you’re spending a program budget to build, your supplier might now simply sell you — better-integrated with the model, and maintained by the people who make it.

First the model, now the company

This is the same story I keep coming back to, escalated one level. First the generalist model beat your specialist model — the general-purpose system outscored the tool built for the narrow job. Now the generalist company is entering your specialist market. The commoditisation didn’t stop at the model layer; it’s moving up, product by product, into the verticals that assumed the model was as far as the labs would come.

The honest limit matters here, though, and the data-honesty rule cuts both ways. Launching a vertical product is not the same as winning the vertical. Claude Science is a beta running on existing models, not a biology breakthrough; Anthropic’s own life-sciences lead frames the drug program around neglected diseases precisely because the obvious commercial targets are already crowded. Entering a market takes a login page. Owning one takes distribution, trust, regulatory approval, and proprietary data — none of which a model vendor gets for free.

What this means

If your business or your project sits above the token layer, the question to keep live is uncomfortable but clarifying: what happens the day my model supplier ships this? If the honest answer is “most of my value disappears,” your value was never really yours — it was rented from the layer below.

The defensible ground is the ground the labs can’t easily take: proprietary data they don’t have, workflows and trust relationships built over years, regulated processes with real approval moats, distribution into places a research lab doesn’t reach. The neglected, the regulated, the data-locked, the deeply embedded — that’s where independents still win, and it’s not a coincidence that it’s exactly the ground a frontier lab finds least attractive. The labs are coming up the stack. The move isn’t to stand on a floor they want. It’s to build where the elevator doesn’t go.

References