Claude just Launched an Equity Research Plugin. Here's Why You Don't Need It.

I just got seriously excited. Christmas Eve level excited.

News reports told me that Anthropic just launched a new set of plug-ins for Claude Cowork specially designed to automate technical knowledge worker roles - and one of them was equity research.

I’ve been waiting for this for some time. As I’ve demonstrated with the stock research reports I post regularly on inferentialinvestor.com, top AI models have the ability with the correct instruction, information, source controls and guidelines to produce institutional level research on earnings events, deep dives, stock initiation reports and more. But the grind has always been that Generative AI models have limitations:

1. They have hard knowledge cut-offs in their training data, often 12-18 months ago which in the case of investing is a recipe for disaster.

2. They cant access gated financial data which is the core of financial decision making. Up to date consensus estimates, EPS revision trends, valuation multiples, filings etc all rely on web search where its usually returning you data from Simply Wall St, or worse, some random Reddit thread.

That means that to run one of Inferential Investor’s deep research workflows I have to source the IR materials manually, extract consensus or other reliable financial data from my data terminal manually, format it and then run the prompt. I’ll admit that gets very tedious when you are cutting across 10+ earnings reports in a day looking for opportunities. There’s too much friction - an aspect I am in development on something to solve for everyone.

So when Anthropic announces they’ve just launched an equity research plugin - I’m thinking, “FINALLY!. It’ll have access to verified data, source materials and skills unavailable to the standard desktop app or web chat interface.”

Unfortunately I was wrong.

It seems for some reason almost blasphemous in the current environment to criticize Anthropic. They’ve gone on an evangelistic PR-driven hyper drive ahead of their planned IPO. Massive feature rollout cadence and even more massive PR effort to get their name in the press every day.

But, the reality of this plug-in is simply very disappointing for the average investor. You see Anthropic’s bread is buttered by enterprises. 80% of their revenue comes from (mainly) large companies. At the same time, they are the most expensive model to run, day to day, out there. I mean orders of magnitude more expensive. As an example, I’ve been testing APIs across the LLMs on equity research workflows. To run my earnings analysis report costs the following via API:

Claude Sonnet 4.6: US$0.52.

Gemini: US$0.12

Deepseek R1: US$0.02

You see the difference. Multiply that by three for a typically comprehensive earnings event workflow: (1) earnings preview to get the financial and contextual expectations set, (2) the earnings analysis itself to see how the company performed versus real expectations (not just revenue and EPS which is only a portion of the picture) and then (3) the earnings call analysis to get the deeper management, Q&A and guidance context for modeling, and you’re looking at approximately $1.50+ for Claude, $0.36 for Gemini and around $0.06-$0.07 for Deepseek. That difference adds up so quickly your bank balance starts to weep.

So I expected that a specialized plug-in, to really solve the pain points of investors, would provide seemless access to data, source materials and better analysis and capabilities that the desktop / web chat interface doesn’t have in order to justify Claude’s super premium pricing. That’s not the case.

First - the underlying models are exactly the same. That means the brain doesn’t change at all. You can access the same brain in your $20/mo PRO subscription.

Second - you require an enterprise account to access the plug-in. PRO and MAX subscribers don’t get it. Why? Well Anthropic charges enterprise users on a $20/mo seat + full API token pricing basis. That means every token gets charged on a variable basis above the seat price. Do more, pay more - pay a LOT more. Get it to access your files and all that information is more input tokens to charge for. Get to pull data from an MCP connection to a database and thats more thousands of input tokens. It builds up quickly.

Third - it doesn’t come with data or source materials bundled at all. Anthropic have partnered with LSEG/Refinitive, S&P Capital IQ and Factset to enable MCP connectors into their databases (so you can query Factset for example with a natural language interface, which is something that Factset actually already has itself) but the user must first have a full subscription to those services that itself runs at USD $2,400 - $10,000 per year. There are cheaper data providers, but Anthropic has not partnered with most of them. Alpha Vantage has a cheaper MCP with Anthropic but its data has some big missing links for what a fundamental investor really needs, Claude cuts off the API data responses with a token limit that makes it useless for consensus retrieval and it doesn’t feed into the plugin as far as I’m aware.

So where has Claude excelled?

Claude equity research reports are very pretty. They have better tool calling to create simple charts and tables and initially I was very impressed with the maturity of Claude’s synthesized conclusions and interpretations in a equity research environment. It seemed more mature, detailed focused and professional and appeared to pick up on softer signals that other models would miss in their conclusions. However, being more than 4x more expensive than Gemini for the same workflow and 25x than Deepseek made me really question whether the value of the output was really that much greater.

So I embarked on a project to see if I could replicate Claude level analysis and interpretation using my workflows in other models. It appears you can (with the exception of charts so far which I have a feeling Gemini will solve this soon). You see, I discovered that most of Claude’s tighter synthesis of analysis comes from what it calls “SKILLS”. Users install the equity research skill (this is distinct from the equity research plugin just to confuse you) via the configuration settings.

Claude Skills & Replicating them in Cheaper Models

This “Skill” provides the model with guidelines to control how it responds for the task, tone, style and even areas it should address in an equity research context. The skills refers the model to reference documents for specific analysis types that might be specialized, such as a DCF valuation. That all sounds fairly useful and sophisticated.

But at the end of the day, the “Skill” is just a fancy name for another system prompt that gets reviewed by Claude before it starts your task. Anyone can replicate this in any model by providing that context in the prompt. You can even access the text of the skill itself (and edit it if you want to) which allows you to see and distill what you need for any other model. It isn’t anything particularly sophisticated and so I summarized what is a long skill document into the aspects I needed, made some critical adjustments to reward weak emerging signal and risk identification and de-emphasize pure historical extrapolation (a noted weakness of all LLMs) and applied it to my standard earnings analysis prompt in cheap old Deepseek.

The result was excellent. Deepseek then started to identify, consider and highlight softer signals that were critical to the conclusions an investor needed to draw when looking at the future implications of the earnings event.

What this does is highlight how important prompt and context engineering is to obtaining really insightful responses. It also highlights how interchangeable these models are. Claude has taken a friction point (long system prompting to guide task specific responses) and made it a feature they can market as a “skill”. But do not be mistaken, it is nothing more than context prompting and you can do it yourself and achieve much the same result in other models.

Claude Plugins - where do they fit?

So where is the value in the plugins in equity research? There is definite value there for certain workflows. Plugins are primarily designed for Cowork to control highly specific task interactions with other apps like Excel, Powerpoint, your files on your computer etc. That means that enterprise customers can prompt Cowork to access their financial database (eg LSEG / Refinitive) retrieve data, integrate that data into a piece of analysis and then update a financial model, investment thesis, equity research report or powerpoint deck with it. Nice - thats the marketing promise anyway. I can see that useful for highly repetitive tasks such as pulling together comparative valuation tables. But the data still needs to be checked as the backbone of this is Generative AI.

Is it a game changer? I’m not convinced. I spent 20 years analyzing stocks and rarely was there a task, other than pure valuation multiple analysis, where it was a plug and play exercise. Even updating financial models was not simply a matter of plugging in the new data points. You had to review the disclosures, adjust for ever changing segment definitions, reconsider your variable and fixed cost drivers and assumptions to work out why actuals differed from your forecasts etc. Doing all that also serves a purpose - you learn more about the company every time. I’d be keen to hear from sell side analysis if they can see a case where they would let a Claude plug-in loose on their financial models. While it has the capability, the reality of it is that what a software engineer sees as programmatic and what an equity research analyst actually does, differ widely. Analysts would end up writing countless natural language prompts to replicate the same work that they do inside their excel models today quite efficiently. It sounds like automation, but in this domain I’m not sure it achieves much. Claude Cowork can access your financial model and extract what it needs for a sell side equity research report. But that can be done anyway without AI and at a far cheaper cost.

Effectively these “plugins” are just more sophisticated “skills” for Cowork to make its file and app manipulation more domain specific. That works in some use cases but solves little for the day to day reality of most others. At some point I think enterprises are going to try all this and then wake up to the realization that Generative AI interfaces aren’t the most efficient ways of getting many quantitative tasks (that also require a lot of experienced human judgement and consideration), done. Use a plugin to get from zero to the structure of an initial financial model fast. It can definitely do that. But then you’ll find yourself going over every cell, assessing and changing the way it builds up the revenue, cost and margin drivers. No one wants to do that via a natural language interface. You’ll be back in excel, grinding away. Very quickly, I think investors will find themselves narrowing what they really use a Cowork plugin for.

Does the Equity Research Plugin Solve Anything for the Average Investor

No. First you need an enterprise plan. Why would any investor open themselves up to variable API pricing when they don’t have to? Second, very few individual investors are going to subscribe to Factset, LSEG or S&P Captial IQ given the cost of those subscriptions. So it doesn’t solve the key data and information access problem for the average investor at all.

Conclusion

My initial excitement has faded to disappointment. There is room for an answer that blends LLM’s contextual analysis capabilities with up to date financial data to synthesize really comprehensive and current analysis. That should be available to all investors, not those than can afford $25k USD/yr seat prices like Rogo, AlphaSense, Bloomberg or Claude Cowork Enterprise + plugin + Factset / LSEG. Anthropic has not tried to solve that. Their path is clear - super premium pricing for enterprises alone.

In contrast, all the R&D I present here at the Inferential Investor is designed to bring the same capabilities in reach of the average investor.

Andy West

The Inferential Investor