Astreus31/40
requests where the Astreus work was preferred
28/40
times the single best file came from Astreus
98%
of files cited the source document for their figures, against 63%
9%
faster to a finished file, on average
The test
We wrote 40 requests the way people actually type them. Some were precise (“Build me a three-statement model for Nahdi Medical”), some casual (“has tesla actually delivered on what elon promised?”). 28 were about Gulf companies and markets, 10 about the US. They asked for Excel models, PowerPoint decks, Word memos and written answers.
Claude ran every request three times with web search only, and three times with web search plus the Astreus MCP server. Nothing else changed: same model, same tools, same wording. That gave six files per request.
Graders saw the six files with every trace of where they came from removed. They ranked them from best to worst and checked the figures against the companies' own reports.
Astreus won 31 of the 40 requests
Each square is one request, coloured by the setup whose files the graders preferred.
Why: the files had more in them
With web search, Claude pieces a company together from news stories, data aggregators and whichever PDFs it can find. With Astreus, it reads the company's own statements, already structured, years deep, with the page each number came from. The difference shows up in what lands in the file.
What the files contained
Share of files in which the graders found each kind of data.
Citations matter most to anyone who has to stand behind a number. 98% of the Astreus files named the filing their figures came from, so a reader can check them in seconds.
Where it went above and beyond
In 16 requests the three Astreus files took the top three places outright. Five of them show what changes when an agent has the real data at hand.
It understood the question
“Ooredoo vs Zain vs STC: which one looks cheapest? Put it in a spreadsheet.”
Web search only
All three files compared mobile phone plans. Claude read “cheapest” as the price of a tariff, and none of the files looked at valuation.With Astreus
All three built a valuation comparison from the companies’ audited statements. The best one rebuilt trailing-twelve-month earnings, adjusted Zain for a one-off item and reconciled its inputs.Kinds of data in the file, of 13 (average of three runs)
A casual question, a researched answer
“has tesla actually delivered on what elon promised?”
Web search only
Scorecards of Elon Musk’s promises with delivery numbers, but no financial history behind them.With Astreus
Each answer checked the promises from Tesla’s events and earnings calls against six to seven years of verified results: profit, free cash flow, capital spending and deliveries.Kinds of data in the file, of 13 (average of three runs)
A memo with nothing missing
“Write a Word investment memo on STC (Saudi Telecom).”
Web search only
Two years of results. Thin on the balance sheet and cash flow, with one memo’s dividend argument contradicted by the company’s own cash-flow statement.With Astreus
Every memo covered 12 of the 13 kinds of data the graders looked for: five years of statements plus the quarters, segments, detail from the notes, guidance and peers, each figure traced to its source.Kinds of data in the file, of 13 (average of three runs)
A deck an analyst can present
“Make me a PowerPoint on DEWA.”
Web search only
Readable overviews. None of the three decks had a balance sheet or a cash-flow statement.With Astreus
Every deck had five years of results and cash flow, a leverage table, the business segments, valuation history and dividend analysis, with each figure cited to the document it came from.Kinds of data in the file, of 13 (average of three runs)
Numbers that tie to the audited accounts
“Build me a three-statement model for Nahdi Medical.”
Web search only
Two of the three models took their figures from a data aggregator and misstated gross profit in every year. One had only three years of history.With Astreus
Every model was built from the audited statements, with page references. In the best one, all 38 checked figures matched the accounts.Checked figures that were correct (average of three runs)
It holds up wherever you ask
Astreus helped most where the open web is thinnest: Gulf companies, whose reports are scattered across exchange sites and PDFs. It still won most US requests, where free data is plentiful. Word memos and decks gained the most; Excel models were the closest contest.
Requests where the Astreus work was preferred
By the format asked for, and by market.
How Claude used it
Given the connector, Claude went straight to it. It looked the company up, pulled the statements and filings, and asked Astreus where a specific figure came from 189 times. It barely needed the web: under one search per run, against about 15 without Astreus.
What it didn't change
Accuracy stayed the same. Claude is careful either way: about 96% of checked figures were correct in both setups (96.4% with Astreus, 96.6% without). The Astreus files carried more figures and more kinds of data at that same accuracy.
The web still won 9 requests. Most asked for data Astreus doesn't hold yet: daily bitcoin and gold prices, the funds buying and selling Microsoft, bond prices for Gulf sukuk, mid-size US banks' property loans, private food-delivery rivals. We are adding to coverage where these point.
Each run used more tokens, because Claude read more real data. It still finished sooner.
How we kept it fair
- The answer keys came first: 1,922 facts taken from the companies' own documents before any file was seen.
- Grading was blind. Names, links and anything that could reveal the setup were removed.
- A second grader re-checked 681 figures independently and agreed on 96% of them.
- The requests were never written to suit Astreus. They are the generic asks an analyst or a newcomer types.
- Every run used Claude Opus 5.5 in Claude Code, in late October 2026: 40 requests, 240 runs.
Give your agent the real data
Astreus connects to Claude and other agents over MCP: financial statements, filings, earnings calls, guidance and macro across the Gulf, the US and Asia, each figure traceable to its source.
Connect Astreus to your agent