Flexreport Finance and Anthropic's Equity Agent Square Off

Published: September 18, 2026

Anthropic made headlines earlier this year when it announced its agents for financial services, promising to revolutionize how research was conducted. As those headlines faded, new ones surfaced, highlighting exorbitant AI costs and broken budgets. Although very impressive, Anthropic's equity agents are just far too costly to be used at scale. And they are also far too slow to even be considered in a real-time setting. We put some numbers around this, running an earnings update for LEN on Flexreport Finance, our own research platform, and Anthropic's equity research agent, running Claude Code's earnings-analysis skill.

On speed the numbers aren't even close. Asked for Lennar, Flexreport handed back the finished report in half a second — 2,500 times faster than the agent's 21 minutes. And that is the easy case: when Flexreport has to build the report rather than serve one it has already built, it takes about 30-90 seconds, still at least 14 times faster.

The two runs

~0.5 s Flexreport, end to end. The report plan was built the moment the earnings event published. The request rendered it — fan out, draw the charts, assemble — with no planning step and no web access.
~21 min Agent, end to end. Started 10:11, finished 10:32. Started from scratch, asking a clarifying question, performing two skill loads, four web searches, three page fetches, at least six calls to Flexreport, then assembling the document.
79.3k tokens pulled by the agent. Context read during the run. The Flexreport side read none at request time: the expensive thinking had already happened when the event landed.
0 vs 7 web requests. Flexreport made no web searches and no fetches. The agent made four searches and three fetches, including a 981 KB transcript page and a 386 KB SEC exhibit.

Orders of magnitude faster, without sacrificing quality

There is usually a trade-off between speed and quality, and you would think that would be especially true for a report created in less than a second compared to a report where an Agent methodically gathers the facts and crafts a rich and detailed report. Yet the quality between the two is very comparable.

Exhibit 1: The two documents, measured

	Flexreport Finance	Claude equity agent
Wall clock	under 1 s	~21 min
Pages	19	21
Words	4,745	4,830
Charts	in-document exhibits	12 interactive
Web searches	0	4
Page fetches	0	3
Transcript	held natively, pinned by section	scraped, 981 KB page

Flexreport's wall clock is the time to return a report it had already built; building one from scratch takes about 30-90 seconds. Word counts from extracted text. The agent's output is an HTML page designed to scroll; page count is what it paginates to at US Letter. Chart counts are figures with an axis, excluding tables.

Anthropic has the wrong mental model for equity research

Anthropic's agents tackle a question like an investigative journalist would: receive a question, gather the facts, write up the report. While this approach has its merits, especially with respect to rigor, it's also extremely time-consuming, costly, and has no way of scaling.

Flexreport takes a markedly different approach. Its agents are stateful, having analyzed companies through time, like a human equity analyst, continuously contextualizing new data, from new SEC filings, earnings transcripts, material news, etc. against what it already knows about a company. This approach has two chief benefits: new datapoints can be assessed with more nuance and better judgement. Secondly, analysis is magnitudes faster and far less costly, since it's not starting from a blank slate. As a result, when you ask Flexreport Finance for an earnings report, it's simply handing you a ready-made one, given its statefulness and having already done the work. Think of it like asking an equity analyst, who's covered American Airlines for ten years, to put together an earnings update for American Airlines versus asking a generalist research analyst.

Same question, two very different calls

On every headline metric the two agree exactly: deliveries of 20,840, new orders of 20,879, home gross margin of 15.8%, diluted EPS of $1.19, and full-year deliveries cut to 80,000 to 81,000. There is no factual disagreement anywhere in the two documents.

Yet they reached opposite conclusions. Flexreport's internal recommendation came out as a SELL. The agent came out HOLD with a $77 price target against a $79.70 spot, cutting FY26 estimates 6% and FY27 20%. While divergence on stock grades is common on The Street, we think Flexreport's grade of SELL is far more appropriate, given the decisively bearish earnings call, price action, and a high rate environment.

Partners, not Competitors

The most interesting thing that emerged from this experiment is that Flexreport and Claude's agent don't have to be mutually exclusive, and it does not have to be an either-or proposition. When Claude's agent was putting together its report, it actually used Flexreport's MCP to bring in historical price and stock grade data to help make its report even better. This is the ultimate goal of Flexreport: seamlessly integrating into your workflow, augmenting it with rich real-time and historical analysis.

Read them both

The two documents are below to serve as a side-by-side comparison. Switch between them and download them in full.

Lennar Q3 FY2026, both write-ups in full

Flexreport Finance

19 pages, returned in under a second — built ahead of the request from a plan that went up when the earnings event published.

Claude equity agent

21 pages, assembled over ~21 minutes from Flexreport data, an SEC 8-K exhibit, a scraped transcript and four web searches.

Methods: both write-ups cover Lennar's fiscal Q3 2026, ended August 31 and reported after the close on September 16, 2026. The Flexreport report was produced by a single call to /get-cached-reports on September 18, which served a cached PDF. The agent run used Claude Code's equity-research earnings-analysis skill on the same date, with timing and tool counts taken from its own execution trace; it had access to Flexreport's MCP server throughout and used it. Word counts are from extracted document text. Recommendations shown are each system's own output and are reproduced to illustrate the comparison. Not investment advice.