Benchmark report · Updated September 2, 2026
Financial Agent Benchmark
Financial answers should survive inspection. On the Vals Finance Agent public set, Drillr MCP scores 94.06%, against 76.30% for Claude Opus 4.6 with built-in web search. In a controlled run over 200 questions, the same model scored 80.76% with Drillr MCP and 33.86% with web search, using 55% fewer tool calls.
01 · The problem
A number without its context.
Search returns passages that resemble the question, without the table, period or filing they came from. That loss of context is why a fluent answer can still be wrong.
A figure is only right for one entity, one line item, one period, one unit and one version of the disclosure. Strip any of those and nothing can check it.
The intermediate value
Claros Mortgage Trust’s EPS note lists several net income figures. Web search stopped at the one before the two-class allocation; the numerator is the one after it.
The superseded value
Pathward’s Q3 2023 net income was corrected two years later in a Form 10-K/A. The original figure is still what news sites and the first 10-Q say.
The keyword collision
Intel’s equity roll-forward has one line for stock issuances and warrants. Similar words appear in separate disclosures with different, larger figures.
02 · The fact record
One fact carries its whole context.
Drillr parses filings into as-reported facts. Each one carries its value and unit, period, filing and filed date, and the page, row and column it was printed in.
The model never reconstructs a period from prose or guesses a unit from a header. Every field comes from the filing that printed the number.
03 · Knowledge graph
Facts are connected, not just stored.
Drillr links facts into a knowledge graph: entities, line items, periods, filings and values are connected, so a query resolves to the right version of a figure, not the most repeated one.
That is what search cannot do. On the benchmark’s Restatement questions, Drillr MCP scored 84.91% against 26.67% for web search.
04 · One call
The fact, with its context, from one request.
filing_search takes a ticker and a question, resolves the company once, and searches the structured as-reported facts. What comes back is the fact with its value, period and filing attached, in one response.
Structured tables sit beside it: run_sql reads financial statements, prices, earnings, analyst, ownership and 8-K event tables directly. On the benchmark that meant 6.91 tool calls per question against 15.39 with web search.
filing_search call.05 · Citations
Every number opens the page it came from.
Every fact carries a citation. In Ask Drillr, the mark opens a card with the quoted passage or table cells, the filing, the period and the page, and a link that opens the filing at that page.
The citation points at a location inside the filing, not a URL. A missing citation is never replaced by a guessed one.
06 · The data underneath
Built on primary filings, market by market.
Primary filings from SEC EDGAR, EDINET and CNINFO, combined with supported market, earnings and alternative datasets.
Coverage is a matrix, not a global promise. Core company data spans all three markets; specialized datasets are narrower, and the interface shows an unavailable state rather than implying support.
Periods and metrics
Preserve the reporting context behind every number.
Fiscal periods
Fiscal year and quarter labels follow the issuer record rather than assuming a calendar-year close. A period is not silently shifted to fit a calendar quarter.
TTM
Trailing-twelve-month values are derived only when enough compatible quarterly observations exist. Otherwise the available reported period is shown, or an unavailable state.
Currency and units
Values retain their reported or listing currency unless a conversion is explicitly labeled. Units are stated so raw values are not mixed across scales.
Adjusted prices
Return calculations that use adjusted close are labeled as such. Market-price timing, corporate-action treatment and source availability can differ by instrument.
Multiple listings
Company routes resolve by ticker. Listings are not merged or compared unless the Company–Security relationship and listing-specific fields can be verified.
Missing data
Missing, partial and unsupported data are separate states. Zero is never substituted for a missing value.