When labs ship plugins, thin software layers lose their premium

Anthropic's Cowork plugins cut double digits from legal and data stocks in one session. Software without proprietary data now trades as exposed.
RELX fell 14%, Wolters Kluwer about 13% and LegalZoom nearly 20% in a single session on 3 February, with data providers close behind. Investors were pricing a threat to a whole category: on 30 January Anthropic had added plugins to Claude Cowork, its tool for handing office work to agents, and open-sourced 11 of them for legal, finance, data, sales, marketing and other functions.
The legal plugin reviews contracts, sorts nondisclosure agreements and tracks compliance. It does not touch the research databases at the core of the incumbents' legal businesses, and the market sold them anyway. That is the signal worth pricing. A workflow that a model can learn from a few text files no longer earns a software multiple, because the company that owns the model can ship it as an add-on to a subscription.
The product itself is modest. Each plugin is a set of files that bundles skills, connectors, slash commands and sub-agents, released to paid Claude users as a research preview. Two days after the sell-off Anthropic released Claude Opus 4.6, which coordinates teams of agents on long professional tasks, and the selling spread to financial data companies. What still earns a premium is what a plugin cannot carry: proprietary data, liability in regulated work and an installed base that switches slowly. Lower public multiples for workflow software lower what strategic buyers pay, and late-stage private rounds follow.
An application built as a prompt layer over a frontier model now competes with its own supplier, as the GPT Store first showed, and the supplier gives the product away. Defensible ground is what the labs do not want to own: data generated by the product itself, integration into systems of record, accountability for outcomes in regulated work, and distribution in languages and markets the labs serve thinly, from Latin America to smaller EU countries. Every founder should be able to say in one sentence why a lab plugin released next month would not take their customers. Pricing should move the same way, from seats toward completed work, because seats are exactly what agents reduce.
Markets overshoot, and one session can mark a mood rather than a repricing. Recovery will separate the two: if incumbents that own proprietary content regain ground faster than pure workflow vendors, investors can price the difference. Adoption is the other gauge, read in how many companies build their own internal plugins once the research preview matures. The decisive fact is whether other labs follow with packaged plugins for the same functions, which would confirm that model providers intend to sell finished work directly to business teams.
Over three weeks the model layer moved up into applications faster than public software prices assumed. The companies that interest Fanaura are those whose value sits in data, accountability and distribution that no model release can copy, the line our venture thesis draws. A thin workflow layer is a feature the labs will ship themselves, and the market has begun to price it that way.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building AI-native products on proprietary data that a lab plugin cannot copy? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.