Anthropic Launches Claude in Excel: AI Financial Analyst Inside Microsoft Spreadsheets

The Part Everyone’s Missing

Financial analysts gain direct API access to institutional data without leaving Excel, accelerating adoption of AI-powered modeling workflows.

Anthropic dropped a new model yesterday. Most coverage will focus on benchmark scores and the fact that it beat its predecessor. That’s not the story.

Here’s what actually happened: Anthropic quietly connected Claude to six of the financial industry’s most important data platforms — S&P Global, LSEG, FactSet, PitchBook, Moody’s, and Daloopa — and made the whole thing work from inside Microsoft Excel. You can now ask an AI to pull earnings data, run comps, and model scenarios across live institutional databases without leaving the spreadsheet where your model lives.

That’s not a model upgrade. That’s a workflow restructuring.

What Computer Use Actually Is — And Why It Took 16 Months

To understand why this matters, you need to understand what Anthropic means by “computer use.” It’s not a chatbot that answers questions about your data. It’s a model that sees your screen — literally, takes screenshots — and operates your mouse and keyboard the way a person would. It can click through a web form, navigate a spreadsheet, switch between browser tabs, and pull information together across multiple open windows.

Anthropic first introduced this capability in October 2024. At launch, they were blunt: it was “still experimental — at times cumbersome and error-prone.” The company expected rapid improvement. On the OSWorld benchmark, which runs AI models through real-world tasks in Chrome, LibreOffice, and VS Code with no special APIs or shortcuts, Anthropic’s Sonnet models have improved steadily across 16 consecutive months of testing.

Sonnet 4.6 is, by a meaningful margin, their strongest computer use model yet. Early users are describing human-level performance on tasks that used to require careful human oversight: navigating a multi-tab research workflow, filling out a complex web form end-to-end, extracting structured data from a messy document.

The analogy worth remembering: the first email clients were also “experimental — at times cumbersome.” Then they weren’t.

Why the Excel Integration Is the Investment Angle

Financial analysts live in three places: Excel, Bloomberg, and whatever document repository their firm uses. Most AI tools require them to leave Excel, go to a chat interface, ask a question, copy the answer, and paste it back. That’s friction. Friction is why AI adoption in enterprise workflows is slower than the hype suggests.

Anthropic’s MCP (Model Context Protocol) connector architecture solves that. If you’ve already set up connectors in Claude.ai, those same connections work in Excel automatically. For analysts with access to FactSet or PitchBook, that means pulling private company valuations, deal comparables, or fund flows directly into a live model — from the spreadsheet — without switching applications.

Daloopa’s CEO Thomas Li put it plainly in his LinkedIn post: “Daloopa’s data plus Claude for Excel equals agentic financial modeling.” That phrase — agentic financial modeling — is worth sitting with. It means the model doesn’t just retrieve data. It reasons across it, builds from it, and handles the multi-step logic that currently eats analyst hours.

Databricks reported that Sonnet 4.6 matches Opus 4.6 performance on OfficeQA, which tests how well a model reads enterprise documents — charts, PDFs, tables — and reasons from the facts it finds. Box tested Sonnet 4.6 on deep reasoning across real enterprise documents and found it outperformed Sonnet 4.5 on heavy reasoning tasks by 15 percentage points.

For a financial services firm, those aren’t benchmark wins. Those are billable hours.

The Price Point That Changes the Math

Here’s where it gets interesting for anyone thinking about enterprise AI deployment. Sonnet 4.6 is priced at $3 per million input tokens and $15 per million output tokens — identical to what Sonnet 4.5 cost. You’re getting a model that approaches Opus-level intelligence, with significantly stronger computer use capabilities, at the same price as the previous mid-tier model.

The historical parallel: this is what happened with cloud computing from 2012 to 2018. Every 18 months, the same compute got cheaper and more capable, which didn’t just make existing workloads cheaper — it made entirely new workloads economically viable. Tasks that weren’t worth automating at $0.10 per query became worth automating at $0.01.

Anthropic also expanded the context window to 1 million tokens in beta. For analysts working with long-form contracts, earnings transcripts, or large regulatory filings, that means the model can hold entire documents — not summaries — in a single request. Asha Sharma, President of CoreAI Product at Microsoft, noted that teams can now work across “large codebases, detailed financial models, and multiple documents without breaking tasks into smaller pieces.”

What Amazon Getting Involved Signals

The same day Anthropic launched Sonnet 4.6, Amazon Web Services CEO Matt Garman announced it was available on Amazon Bedrock. That’s not a coincidence — it’s a coordinated go-to-market.

Amazon’s cloud infrastructure gives Anthropic enterprise distribution at a scale it couldn’t achieve alone. And AWS needs Anthropic models to stay competitive with Microsoft Azure, which has OpenAI, and Google Cloud, which has Gemini. The three-way cloud competition means that every major Anthropic model release now arrives with immediate enterprise availability across all three major platforms.

For investors tracking the AI infrastructure trade, the pattern is consistent: model launches are no longer just research announcements. They’re distribution events. NVIDIA chips train the model, AWS, Azure, and Google Cloud distribute it, and companies like FactSet, PitchBook, and LSEG build integrations on top. Every layer has a margin story.

The Security Problem Nobody Talks About at Launch

Computer use comes with a real risk. When a model can operate a browser and navigate websites, malicious actors can attempt to hijack it by hiding instructions inside the page content — what’s known as a prompt injection attack. A compromised AI agent that has access to your financial data, your spreadsheets, and your browser sessions is a meaningful security concern.

Anthropic disclosed in Sonnet 4.6’s system card that the model shows “major improvement compared to its predecessor” on prompt injection resistance, performing similarly to the more expensive Opus 4.6. This is the kind of safety improvement that rarely gets headline coverage, but it’s the condition that makes enterprise deployment actually viable.

No CFO is going to approve an AI with access to FactSet data and Excel models until they’re confident it can’t be hijacked by a malicious document or website. The security improvement matters.

What to Watch

The number to track is enterprise adoption velocity on MCP connectors. Anthropic now has six major financial data integrations live: S&P Global, LSEG, FactSet, PitchBook, Moody’s, and Daloopa. Watch for whether those six expand to a dozen by Q3 2026. If major accounting software — Workday, Oracle Financials, SAP — adds MCP connector support, that’s the signal that computer use has crossed from “interesting demo” to “enterprise standard.”

The second thing to watch: whether the $3/$15 pricing holds at 1 million token context. Longer context at the same price means more complex workloads become economically viable. If Anthropic prices 1M context at a premium when it exits beta, it slows adoption. If it stays flat, the financial services use case accelerates significantly faster than the market expects.

The AI analyst isn’t replacing financial analysts. It’s changing what analysts spend their hours doing. The firms that figure out the workflow first — and there are already early adopters at Cursor, Replit, GitHub Copilot, and Harvey running this in production — will have a measurable productivity advantage over those who don’t.

That’s not a prediction. The customer quotes in Anthropic’s own launch announcement make it plain.

—