Blog · 2026-10-22 · Ashutosh Agarwal

Matterfact vs. Claude: What investment teams get beyond the model

Matterfact runs on Claude and builds the research desk around it: agent swarms that work around the clock, methods for every sector, answers that carry their evidence, and the fund's own data inside its own tenant.

TL;DR

• The same Claude model does better work inside Matterfact. Claude Opus 5.5 scored 75.6% on Vals AI's finance tasks in our agent, against 58.6% in Vals AI's own harness, and ran 19 times faster at a fifteenth of the cost.
• Agent swarms take on long-horizon research and keep going when no one is prompting them. Three swarms ran for three to seven weeks each and made 224 dated, testable predictions.
• Earnings and news answers arrive with the evidence attached. Smart Highlight answers in management's own words; read-throughs show how each linked stock took the story.
• Every company gets the method its sector requires: 5,000+ skills across 163 GICS sub-industries, 400+ playbooks and a shared application around each thesis.
• The fund's own data, and data no vendor sells, sit around every answer: one shared layer over its warehouse, files and email, a Snowflake semantic layer and 120M+ podcast episodes, in a single-tenant deployment with role-based access down to the file.

We hear a version of the same question in almost every first conversation with a fund:

"Our analysts already use Claude. Why would we need Matterfact?"

The question gets sharper every quarter. Claude now connects to licensed data such as FactSet and S&P Capital IQ, ships finance plugins for equity research and private equity, and works inside Excel and PowerPoint. For an individual analyst, it is an excellent assistant.

Our answer has not changed. We think Claude is the Rolls-Royce engine for investment research, which is why Matterfact runs on it. But a fund cannot run its research on an engine alone. It needs the aircraft around it: agents that work every name around the clock, methods for every sector, answers that carry their evidence, and the fund's own data, all inside the fund's own tenant. In our tests, the same Claude model was 17 points more accurate inside Matterfact than in Vals AI's own harness.

This article sets out the eight things Matterfact adds around the model: first in one table (Exhibit 1), then one section each.

heading: What we cover
The quick answer: eight things Matterfact adds around Claude
1. The same Claude model does better work inside Matterfact
2. Agent swarms keep researching when no one is prompting them
3. Earnings and news answers arrive with the evidence attached
4. Every company gets the method its sector requires
5. The fund's knowledge belongs to the fund, not to one analyst
6. Analysts can question the fund's own warehouse in plain English
7. Matterfact adds data that no vendor sells
8. The fund's data stays inside its own tenant
Use Claude for general work and Matterfact to run research

The quick answer: eight things Matterfact adds around Claude

Claude covers more of a fund's general work every quarter. What it does not provide is the research desk around the model, which a fund would otherwise have to build, staff and maintain itself (Exhibit 1).

Exhibit 1
Claude now reaches licensed data and Office files; the research desk around the model is what Matterfact adds.

Note: Claude capabilities are taken from Anthropic's public announcements through October 2026, for Claude Enterprise with its finance connectors and plugins. Both products change quickly, and we update this comparison when they do.
Source: Anthropic; Matterfact

1. The same Claude model does better work inside Matterfact

The software around a model, its tools, instructions and data, decides how much of the model's ability reaches the answer. When we ran Vals AI's public Finance Agent tasks inside our agent, Claude Opus 5.5 was 17 points more accurate than in Vals AI's own harness, 19 times faster and 15 times cheaper per task (Exhibit 2).

Exhibit 2
Inside Matterfact, Claude Opus 5.5 was 17 points more accurate, 19 times faster and 15 times cheaper per task.

Note: Vals AI's figures are on its 450 private questions; ours are on the 27 public ones, graded with the benchmark's own rubrics. The full method is in our eval report.
Source: Vals AI Finance Agent v2 leaderboard, October 12, 2026; Matterfact analysis

The effect is not specific to Claude. All four frontier models we tested scored 7 to 17 points higher in our agent. A harness change can be worth more than several model releases: on Vals AI, four Gemini Flash releases gained 3.6 points, while moving Gemini 3.8 Flash into our agent gained 7.1.

The grading needed a practitioner too. A buy-side investor reviewed every rubric item and found some that were wrong and others that were judgment calls, so we put the best achievable score at about 83%, not 100%.

Matterfact also does not tie a fund to one model family, and it owns part of the model layer itself. Our own fine-tuned router sends each task to the model that does it best, and we keep tuning it on how its choices perform:

• Claude for long-horizon research, such as a revenue bridge or a full initiation.
• OpenAI for extraction and for writing a coherent investment memo.
• Google Gemini for long-context work and web search.
• Small models we fine-tune ourselves, running inside your tenant, for narrow, high-volume classification.

For a fund, this is insurance as much as performance. When the leaderboard moves, the routing moves with it, and nobody re-platforms.

2. Agent swarms keep researching when no one is prompting them

Agent swarms are the frontier of AI in finance, built for long-horizon research that spans days to weeks. Give one a research goal, and dozens of agents drawn from 20 specialist types divide the work, challenge the thesis every cycle and keep going until you stop them. Claude's agents work for one user, on tasks that user sets; a standing desk of agents that researches a name for weeks and keeps score is something a fund would have to build on Anthropic's developer platform.

On 1 September we gave a swarm one job: work out what Meta is worth. Over the next seven weeks, it:

• flagged about $356B of data-center leases signed but not yet started, which became one of the two biggest items when it revalued Meta after its results;
• kept a New Mexico court case on its list after nearly dismissing it twice; that case ended in a $25B penalty order;
• predicted that the results-day move would be large and would go against the earnings beat. Meta beat and fell 7%.

Across Meta, Hims & Hers and Borr Drilling, three swarms made 224 dated predictions, wrote 221 questions for management and industry experts, and catalogued more than 1,600 sources and research notes. Every name on a coverage list can get that depth, and analysts keep the work machines can't do: meeting management, running expert calls and deciding. The three case studies are in Introducing Matterfact Agent Swarms.

Swarm figures are research outputs, shown to illustrate how the system works. They are not investment advice.

3. Earnings and news answers arrive with the evidence attached

Ask a chat assistant about an earnings call and it writes a summary, which an analyst then checks line by line, because one invented quote in a note to the PM costs more than the summary saved. Smart Highlight never writes a word. It answers a plain-English question with the sentences management actually said, so there is nothing to hallucinate.

It searches by meaning, not keyword. On our sample call, a search for "capex" finds one sentence, and it is the analyst asking the question. Asked how much the company is spending on AI and capex, Smart Highlight finds five answers, none of which uses the word. The same question runs across quarters and competitors. For an analyst with 25 names and 25 peers, that turns 200 calls a year, about five working weeks of reading, into a set of questions. More in Introducing Smart Highlight for Earnings Calls.

News has the opposite problem: the story that moves a name is often about another company. Feeds filtered by ticker miss a customer's capex plan or a supplier's capacity, and more than a decade of academic research finds that linked stocks react to such news with a lag. Read-throughs in Matterfact News read the news on every competitor, supplier and customer of the names you cover, judge each story separately for each name, and show how the stock has traded since (Exhibit 3).

Exhibit 3
One story about a customer reaches three of the names an analyst covers, each with its move since the story.

Note: Example stories and price moves, for illustration.
Source: Matterfact News

4. Every company gets the method its sector requires

Great investors do not apply one model to every company. The right analysis of a bank is not the right analysis of a biotech, a REIT or a chipmaker: the metrics differ, their definitions differ, and so do the comparisons that mislead. Claude's finance plugins are organized by job, such as equity research or private equity. Matterfact's are organized by sector.

• Skills encode the method. More than 5,000 skills, each a tested way of doing one piece of analyst work, cover all 163 GICS sub-industries. Our analysts and data scientists tune each one by hand, because a software analyst's earnings review is not a bank analyst's.
• Playbooks chain them into whole pieces of work. More than 400 one-click playbooks run from the 6:30 a.m. brief to a full initiation, on demand or on a schedule.
• Both are shared and permissioned. When a senior analyst refines how the firm reviews a print, that becomes how everyone's agent reviews a print, the same afternoon. Firm and team skills are managed without code, and run the same way as ours.

The method then lives in an application built around the thesis: an earnings-season tracker for a coverage list, a live sum-of-the-parts model, or a dashboard of the few indicators that would change the team's view. It brings internal data, purchased alternative data and web evidence into one place, because a thesis is easier to test when the evidence sits together (Exhibit 4).

Exhibit 4
A thesis application starts from the questions that decide the investment, and stays current and shared.

label: A page vs. an application
tone: verdict
left: A Matterfact thesis application
right: A generic page or dashboard
Starts from the questions that determine a specific thesis :: Starts from a generic workflow
Combines live feeds, web evidence and internal systems :: Holds the data pasted into it
Stays current as new evidence arrives :: Refreshes what it was given
Permissioned and shared with the investment team :: No shared working context
Versioned as the team's thesis evolves :: Rebuilt when someone wants changes

Source: Matterfact

5. The fund's knowledge belongs to the fund, not to one analyst

Claude's memory is scoped to a project, and what one analyst builds tends to stay in that analyst's account. When the best analyst leaves, their way of working leaves too. Matterfact brings the fund's data into one AI-ready layer that every agent, skill and colleague shares, and keeps each source current on its own schedule (Exhibit 5).

Exhibit 5
Every source the fund already has stays current in one shared layer, alongside licensed data.

Source: Matterfact

On that layer sit the fund's sell-side research, internal notes, IC memos, expert-call transcripts, models and diligence files, and every skill and application draws on it. Several colleagues can work with the agent in the same chat, and role-based access, down to the file, keeps each answer within what that person is entitled to see.

• A new analyst inherits the fund's context on day one, instead of rebuilding it over eighteen months.
• "What did we conclude about this name last year, and why?" comes back with the underlying documents attached.
• Analysts covering adjacent sectors stop duplicating work without knowing it.

6. Analysts can question the fund's own warehouse in plain English

A connector lets a model reach a warehouse; it does not tell the model what the tables mean. A column called AMT_NET or BOOK_CD carries the fund's own definition, and without it every answer rests on a guess. Claude can query a warehouse through a connector, but documenting what the tables mean is left to the fund. Matterfact builds that map itself, as a Snowflake semantic layer that IT governs and every agent shares (Exhibit 6).

Exhibit 6
The agent builds the map of the fund's warehouse, then refines it with every cited query.

variant: chain
numbered: true
label: Building the Snowflake semantic layer
Read the schema :: Tables, columns and types
Fill in missing descriptions :: One line per table and column
Import semantic views :: The fund's metrics, dimensions and joins
Learn from use :: Each cited query refines the map

Source: Matterfact

With the map in place, an analyst asks "What is our net exposure to semiconductors?" and gets an answer built on read-only SQL, with the query cited. A rules file keeps every agent to the fund's preferences, such as "For consensus estimates, always use FactSet" and "Never use web sources for reported financials". Matterfact also writes every dataset it builds back to the warehouse as governed tables, under the roles and grants the fund already administers.

7. Matterfact adds data that no vendor sells

Claude's connectors reach the vendors any fund can license, and Matterfact reads licensed sources too, such as FactSet and Quartr. The evidence that moves a thesis, though, often sits in data nobody packaged. Matterfact adds two such sources.

• Podcasts, at 120M+ episodes. Full transcripts, tied to tickers, of where CEOs, CFOs and investors speak at length and with less compliance filtering than on an earnings call. See podcast search.
• Any website, as a dataset. The agentic web scraper turns sources that were never sold as data into structured datasets that refresh on a schedule: every VIN listed on Carvana as a used-car sales proxy, Phase III enrollment on ClinicalTrials.gov, every Chipotle opening. It reads sources in their own language, from Japanese job boards to China's weekly EV insurance registrations, and maps them to tickers.

Both arrive as feeds, kept current at full history, not as files trimmed to fit a chat upload. Changes route into alerts and into the applications the team uses to test its view. The work is fast: one fund used the data engineer agent to map three retail chains' store footprints, work it had budgeted one to two months for.

8. The fund's data stays inside its own tenant

Claude Enterprise is a shared cloud service with enterprise controls, and Anthropic does not train on its customers' data by default. Matterfact goes further, with controls vetted by CISOs at top-tier hedge funds:

• Single-tenant. One isolated deployment per fund, on AWS or Google Cloud, or inside the fund's own VPC or on premises.
• Audited. SOC 2 Type II, with third-party penetration tests.
• Zero data retention. Model providers keep nothing, and the fund's data trains no one else's model.
• Traceable. Role-based access through the fund's identity provider, down to the file, and an audit trail of every query, source and output, exportable to its SIEM.

The details are on our security page. Matterfact is the firm-wide AI system at one of the world's top four hedge funds, used on every desk, and its Head of AI puts the outcome in one line: "They shipped in weeks what would have taken us quarters."

Use Claude for general work and Matterfact to run research

Claude and ChatGPT are core applications for any modern firm. Your team should keep using them for what they do best: writing, brainstorming, coding, learning, and the hundred questions that do not need a fund's research infrastructure. Choosing Matterfact is not choosing against them. Matterfact runs on frontier models, including Claude, and Claude can call Matterfact in turn: every answer and dataset is reachable through our API and MCP server, from Claude, ChatGPT or the fund's own tools. Our Matterfact connector for Claude already brings our podcast intelligence into a Claude chat.

For a fund evaluating both, four questions separate a research platform from an assistant:

• Does it keep working on a name when no one is prompting it?
• Can every number and quote be traced to the filing, call or page it came from?
• Does it know the method for each sector, what the fund has already concluded, and what the fund's own tables mean?
• Does the fund's data stay inside the fund's own tenant?

The one-line summary: Use Claude and ChatGPT for general work. Use Matterfact to run investment research.

FAQ

Does Matterfact replace Claude?

No. Matterfact is built on frontier models, including Claude. It adds the layer a fund needs around the model: agent swarms, sector skills and playbooks, the fund's own knowledge and data, earnings and news answers with their evidence attached, and orchestration across models. It also works from inside Claude: our API and MCP server make every Matterfact answer and dataset callable from a Claude chat.

Claude now connects to FactSet and S&P Capital IQ. What does Matterfact add?

Those connectors put licensed data into a chat, and they are useful. Matterfact reads licensed sources too, and adds what a research desk needs around them: research that runs without prompting, answers in management's own words, data no vendor sells, and the fund's internal systems, all in a single-tenant deployment.

Which models does Matterfact run on?

Claude for long-horizon research, OpenAI for extraction and memo writing, Gemini for long-context work and web search, and fine-tuned small models inside your tenant for high-volume classification. A fine-tuned router picks the model for each task, and we re-evaluate its choices as model quality changes.

Where does our data live?

In a single-tenant deployment: your fund's own isolated environment, on AWS or Google Cloud, or inside your own VPC. Your data is never pooled with other funds and never trains anyone else's model. The details are on our security page.

We already built some Claude connectors ourselves. Is that wasted work?

No. Matterfact speaks the same standard (MCP), so the fund's own MCP servers connect live, alongside Snowflake, SharePoint, OneDrive, Outlook and SFTP.

Can we start small?

Yes. Many funds start with a single team, a single workflow or one swarm on a name they cover, and expand from there once the value is obvious.