Claude Personal Finance: What It Reads, and What It Can’t Do

Claude Personal Finance is a read-only beta from Anthropic, launched November 5, 2026, that lets paid US subscribers link bank and card accounts through Plaid and ask Claude about their money. It reads balances and up to 24 months of transactions, and it can't move, manage, or invest anything.

As of November 6, 2026, we have not tested the beta, which is rolling out gradually to US accounts. Everything below comes from Anthropic's announcement, the Claude Help Center, and OpenAI's documentation, and we mark our own analysis as such. Valletta builds client software with Claude and other models, so weigh our take with that in mind.

What is Claude Personal Finance?

Claude Personal Finance is a dedicated page in Claude's side menu plus a Plaid connection that also works in regular chats. The page shows widgets for spending, cash flow, accounts, subscriptions, upcoming bills, and short insights, which you can edit and reorder. You can then ask questions in plain language, on that page or in any chat.

The launch follows Claude for Legal and Claude for Financial Services, but those target firms. This one is for your own checking account. Anthropic calls it a first step and says it wants Claude to work alongside licensed professionals and regulated firms on bigger financial decisions, and it is inviting banks and wealth managers to help shape that.

What can Claude see in your bank accounts?

Claude sees only the accounts you choose to share in Plaid, and only the data types below. Plaid's own consent screens may mention investment, credit, and loan data, so the screen can read broader than what Claude actually uses today.

Plaid passes only the last few digits, but some banks put a whole account number into an account name or a transaction description, and then Claude can see it there.

What can you ask Claude about your money?

The questions Anthropic highlights are about understanding what already happened, along with what is due next:

• Where did my money go last month, by category, and is that higher than usual?
• How much do I have left before payday, and which bills are still due?
• Are there duplicate charges, unexpected fees, or bills that have crept up?
• Which subscriptions am I paying for?
• What is APR, and how much interest did I pay over the past year?
• Can you build a chart, bill calendar, or dashboard from this?

Where the numbers can be wrong

Anthropic tells users to check important figures against their bank, and the help article lists three specific failure modes:

• Stale balances. A balance can be a day or two old.
• Double counting. An account linked through two logins can be counted twice if the bank describes it differently under each.
• Silent gaps. While Claude can't read a bank, the page's spending figures leave that bank out, so a budget can show more money left than you have.

The third one matters most. Banks sometimes end Claude's access, and a budget that quietly drops a bank looks healthier than it is. If a number looks good, check the linked accounts list for a "Sign-in expired" status first.

How do you set up Claude Personal Finance?

1. Open Claude on the web, Claude Desktop, or the mobile app, and select Personal Finance in the side menu. It may sit under More.
2. Select Get started, then choose your goals or skip them.
3. Select Link accounts with Plaid and review what Claude will be able to access.
4. In Plaid, pick your bank, sign in, and choose the accounts to share. Finish within 10 minutes.
5. Back in Claude, wait for your accounts to appear. Transactions can take a few minutes to sync.

If the menu item is missing, check that you are on a personal paid account, not Team or Enterprise, and not on a VPN that exits outside the US. It may also just not have reached your account yet.

Is it safe to connect your bank account to Claude?

Mostly yes for the connection itself, with real risks in what happens to the data afterward. You sign in through Plaid, Claude never receives your password, and its access is read-only, so it can't move money. The remaining exposure is in chat history, shared links, and artifacts.

The password point is topical. In late October 2026, Bank of America emailed customers warning them not to give banking credentials to AI agents and saying they are responsible for activity that results from that access. The report doesn't say whether the email covers Plaid-style connections, though it describes bank-authorized read-only links as an option that doesn't let an assistant move money.

What to watch once your accounts are linked:

• Chat history. Bank data appears in your conversation and stays until you delete the chat. Deleted chats leave Anthropic's back-end within 30 days, with longer retention for flagged content, feedback, or legal reasons.
• Shared links and artifacts. A shared chat link shows any figures from your accounts, and figures saved in an artifact are shown to everyone who opens it. An artifact also stays after you delete the chat that made it.
• Chats outside the hub. The no-training promise covers the Personal Finance hub. Elsewhere, your privacy settings apply.
• Default access. Claude reads linked accounts in chats without asking each time. You can switch Plaid off for chats on any platform, and on mobile you can also set it to Needs approval.
• Lapsed plans. When a paid plan ends, your banks stay linked until you unlink them. Unlink before you cancel, or disconnect at my.plaid.com.

Prompt injection: the risk the docs don't discuss

This part is our analysis, and we haven't tested it. Prompt injection is when a model treats text inside the data it reads as instructions. Your bank's transaction descriptions are partly controlled by third parties, because merchants choose what appears on the statement, and Claude reads that text. We didn't find this addressed in the launch post or the help article.

Some product limits would reduce the damage. Linked accounts can't be used in scheduled tasks, Cowork, or Claude Code, a chat that already ran code can't use your accounts, and Claude asks before publishing an artifact from a chat that used them. Anthropic doesn't say these limits exist for that reason, but they look like containment to us, and they are a sensible template if you are building something similar.

Is Plaid safe?

Plaid is a widely used connection layer with recognized security certifications, but how much of your login it handles depends on your bank. Whether Plaid stores your username and password comes down to the connection type, and this applies to every app that uses it, Claude and ChatGPT included.

Source: Plaid Consumer Help Center.

You can only partly tell which one you have. If you are redirected to your bank's own website, that is OAuth and Plaid doesn't get your login. If you type your login into Plaid's screen, it may be either of the other two. Plaid says it never shares your credentials with the connected app in any case.

• Certifications. OpenAI's help page describes Plaid as certified to ISO 27001 and ISO 27701 and SOC 2 compliant.
• Past lawsuit. Plaid agreed to a $58 million class action settlement in 2021 over claims that it collected bank logins and sold transaction data without adequate consent. A settlement is not a finding that those claims were true. The terms included deleting some data, clearer disclosures, and storing less.
• Selling data. Plaid's consumer privacy notice says it doesn't share your data for its own marketing, and that it shares with outside companies for their marketing only with your consent or at your direction.
• Control. You can see and revoke connections, or request deletion, at my.plaid.com.

Can Claude replace a financial advisor?

No, and Anthropic doesn't claim it can. Claude Personal Finance isn't designed to give investment advice and can't see holdings, so it can't assess a portfolio, and Anthropic says bigger decisions are where it wants to work alongside licensed professionals.

The best recent evidence on AI financial advisors in general is mixed. A study from MIT Sloan and Stanford researchers had 1,000 adults ask GPT-5.2, GPT-5.6, or Gemini 3 Flash for advice in their own words, then simulated following it over a lifetime. The advice was generally sound on saving and diversification, but it adjusted poorly to shocks such as losing a job. Prompting skill also mattered: users with no prior experience of AI financial advice ended up about $100,000 (6%) poorer at age 60 than experienced users. The study didn't test Claude, and it measured advice, not spending analysis.

Claude vs ChatGPT for personal finance: which is better?

ChatGPT is broader. Claude is narrower, and stricter on memory and hub-chat training, while OpenAI states a deletion window for synced data. Neither can move money. OpenAI launched its Plaid-based feature on May 15 for Pro users, then widened it, and the current help page lists Free, Go, Plus, and Pro (see how those ChatGPT tiers compare).

Sources: Claude Help Center and OpenAI Help Center, both checked November 6, 2026. Early coverage of ChatGPT describes it as Pro-only; that is out of date.

If you want one view of spending, investments, and credit, ChatGPT is the more complete tool today. If your question is mainly "where did my money go" and you would rather the assistant not remember your finances, Claude's narrower design is the better fit. OpenAI's help page also states a 30-day deletion window for synced data. Claude's help article and linked retention page don't give one for synced data, though deleted chats leave Anthropic's back-end within 30 days.

What is Anthony Pompliano accusing Anthropic of?

Anthony Pompliano, CEO and chairman of the rival app Silvia, alleges that Anthropic staff cultivated Silvia's engineers to learn technical details and plans before launching a competing product. None of it is verified. He made the claims in a thread on X within hours of the launch, and it passed one million views:

• Starting in April, Anthropic staff courted Silvia's engineers, citing interest in the company's growth and compute usage, while actually seeking corporate secrets, technical details, and future plans.
• This included a steak dinner that someone said was on Anthropic CEO Dario Amodei.
• He says Silvia's team refused to share sensitive information.
• He says Anthropic wanted the information so it could build a competing product.
• Anthropic staff messaged Silvia days before launch to warn that a competing product was coming.

He called the result "a Claude wrapper that is masquerading as a personal finance product." He also said later in the thread that he isn't angry about it and that this kind of thing is old in business.

What is and isn't established

The thread names no Anthropic employee beyond the Amodei reference and includes no documents. It doesn't say who made the approaches or what was discussed. FourWeekMBA says nothing it read shows Claude's feature drew on Silvia, and neither did anything we read. FourWeekMBA reported no public Anthropic response as of 00:34 UTC on November 6, and we found none later that day. Decrypt also covered the allegations. In our view, a model vendor talking to a high-usage customer about compute would be ordinary account management, and Pompliano says that was the stated reason. The filings we read don't say Silvia was an Anthropic customer, so the claim depends on evidence that isn't in the thread. Reaction on X was split: many replies mocked the idea that a frontier lab needed a small startup's secrets, while many quote posts sided with Pompliano.

One thing is on the record. Silvia's October 2026 filings list foundation model providers favoring their own competing products among its risks. Its February 10-K, filed while the company was still ProCap Financial, used the same phrase about cloud providers such as AWS, Azure, and Google Cloud, not model vendors. So the general risk was disclosed months earlier, and it was extended to model providers about four weeks before this launch.

What Silvia's benchmark shows

Pompliano backed his claim that Claude is the inferior product with a chart. Claude comes second in all three tests:

Scores read from the chart Pompliano posted. Its footnote labels them internal benchmarks and not independent rankings.

Three caveats apply before reading this as a win for anyone. Silvia designed and ran the benchmarks itself, though it publishes them on Hugging Face. The results came out in August and September, before Personal Finance shipped, so they can't measure it. And the tax dataset has 200 questions: 100 adversarial ones, such as California's treatment of qualified small business stock, and 100 planning scenarios with synthetic personas. Silvia's August release names the 8.40 entry as Claude Desktop. That tests tax guidance on hypothetical cases, whereas Claude Personal Finance answers questions about your own transactions.

Pompliano's claim that compute costs are about 97% lower on Silvia's own model is also a company figure. Silvia's September release applies 97% to one specialist model and puts the drop in overall cost per question at 59%, and FourWeekMBA says the 97% isn't in the filings it read.

The products overlap on spending, but less elsewhere. Silvia's filings describe connections to brokerage, retirement, crypto, real estate, and alternative investment accounts, and its site also shows bank data and transactions and talks about recommendations, while its press release says it doesn't provide investment advice. Claude reads no holdings and says it isn't designed for investment advice. Anthropic's stated roadmap could move it closer to Silvia's territory.

Can AI labs compete with the startups built on their models?

Yes. A model provider can build into its customers' markets, and Silvia's own filings now list that risk. Indie founder Pieter Levels predicted on X that the labs will work through industries one at a time. If your product is a vertical layer on a model, here is what we would do:

1. Keep the model swappable. Route calls through one layer so you can change vendor or run your own. Our self-hosted LLM cost breakdown shows when owning inference pays off, and the current frontier model comparison shows how quickly the leaderboard moves.
2. Own your evaluations. Silvia's public benchmark is contestable, but having one is the right instinct. Tests built on your real cases tell you whether a new model or vendor actually helps.
3. Compete on what the lab won't touch. Claude's limits are visible: no holdings, no rates, no regulated advice, US only. Licensed workflows, proprietary data, and integrations are harder to copy with a sidebar item. We built this kind of narrow system for an advisory firm generating LLM financial reports.
4. Copy the permission model. Read-only scopes, no password handling, no memory on sensitive chats, and isolation from code-running tools are cheap to design in and expensive to add later.

FAQ

Is Claude Personal Finance free?

No. It requires a paid US plan: Pro ($17 a month billed annually or $20 monthly), Max ($100), or Ultra ($200). Free, Team, and Enterprise accounts don't get it. ChatGPT Finances does reach free US users.

Is Claude Personal Finance available outside the US?

No. Only US banks can be linked, and the feature doesn't appear when you are outside the US or on a VPN that connects from another country. Banks you already linked stay linked until you return.

Can Claude move money from my bank account?

No. Access is read-only, so Claude can't transfer funds, pay bills, make trades, or manage accounts. It does read your linked accounts in chats without asking each time, unless you change the Plaid setting.

Does Anthropic train its models on my financial data?

Conversations in the Personal Finance hub aren't used for training. Anything you share with Claude outside the hub follows your normal privacy settings, so ask money questions inside the hub if that matters to you.

What happens when I unlink a bank from Claude?

Claude stops reading every account tied to that bank login. You can't unlink a single account. Anything Claude already said stays in your chats, artifacts, and shared links until you remove it. Unlinking your last bank also deletes your dashboard, budgets, and setup answers, and that can't be undone.

Is Claude or ChatGPT better for personal finance?

ChatGPT covers more: it is on free plans, shows investment holdings, and adds credit scores. Claude covers less but keeps chats that used your accounts out of memory, and doesn't train on hub chats. OpenAI is clearer on deleting synced data. Choose based on whether you need portfolio data.

Is Claude an AI financial advisor?

No. It explains your spending and terms like APR, and calculates things such as interest paid. Anthropic says it isn't designed to give investment advice, and it can't see your holdings anyway. For large decisions, Anthropic points to future work with licensed professionals.

What is CFO Silvia?

CFO Silvia, now branded Silvia, is a personal AI finance app for investors run by Anthony Pompliano. It connects bank, brokerage, retirement, crypto, and real estate data. Its parent company, formerly ProCap Financial, renamed itself Silvia, Inc. in September 2026 and trades on Nasdaq as SVIA.

Did Anthropic copy Silvia?

That is not established. Pompliano alleges Anthropic staff fished for technical details, but his thread names no employee and shows no documents, and by his own account his team refused to share sensitive information. Nothing we read shows Claude's feature used anything from Silvia. We found no public response from Anthropic as of November 6, 2026.