Ten days or so after payday, the same question comes back every month: where did this month's money actually go? Anthropic put the job of adding it up across three or four banking apps into Claude on November 5. Personal Finance connects bank and card accounts and answers in everyday language from balances and transactions — and the first line of its design is the boundary of its authority. Access is read-only, so Claude cannot move, manage, or invest your money, and the ticket opens in the US from the $20/month Pro plan. OpenAI has already widened the same capability down to its free tier, so this article separates, from the official materials, where the latecomer's design splits.
The short version
A read-only bank connection — linking happens only when you sign in to your bank yourself through Plaid, Claude never sees your bank password, and transactions on chosen accounts are read back up to two years (official announcement and help center, checked directly)
Investments, loans, and cards are read only down to balances and credit limits, not holdings or rates — not an AI that knows all your assets, but one that knows your cash flow
A paid US beta from $20/month against a ChatGPT that widened to free — with Claude answering with no ads and hub conversations excluded from training as its defaults
The connection follows the path three companies share. Only when a user signs in to their own bank on Plaid's screens does the link form, and Claude neither asks for nor sees the bank password. What Claude then sees, for chosen accounts only, is the bank's name, each account's name, type, and last four digits, balances and credit limits, and up to two years of transactions. Up to 16 banks can be linked.
The scope needs cutting once more, source by source. The "up to two years of transactions" line is what the announcement writes about chosen accounts in general; the help center then limits the kinds of data by account type. For investment, loan, and card accounts, it states, Claude reads balances and credit limits — not holdings, trades, rates, or terms. Even if Plaid's consent screens ask to share more, what Claude actually reads today is only that list. This is closer to an AI that knows your cash flow than one that knows all your assets, and it is also why the example question — what APR is, and what interest cost you this year — works: it runs on the card-balance and transaction side.
The list of things it can do is built from cash-flow questions: summing this month's spending by category, comparing it with previous months, flagging double charges, unexpected fees, and bills that have crept up over a year, and counting the payments due before payday to compute what is left. It can build a weekly spending chart, a payment calendar, and a dashboard of numbers you choose, and keep adjusting them. How accurate any of this is in practice has no record outside the company yet. The announcement itself writes that Claude can make mistakes, so ask which transactions sit behind a number and check important figures with your bank. The line that it is not designed to give investment advice sits in the same place.
The only case that can be checked today is a demo the company built. Late in the month, with payday ten days away, a person asks where this month's money went; spending totals $4,309.37, more than half of it rent and utilities, $1,024.51 of it food and drink; follow-up questions surface a membership renewal, and after four payments due before payday, $1,393.22 is left. The more concrete the numbers, the clearer the caution. These amounts are figures from a demo example in Anthropic's announcement, not verified results from a named user's account. On the earlier ChatGPT side, the official page at least carries a named tester's words — Richard K. Sohn, saying the connection helped set a realistic monthly plan to pay off a mortgage — while the Claude side has no counterpart named user experience in its official materials yet. Even allowing for a beta in its first week, the thickness of cases is where Claude trails on this axis.
For the same-tier capability of reading bank accounts, the prices and scopes the official pages settle today split like this.
| Capability (bank linking, read-only) | Price (official pages, checked directly) | Status and reading scope |
|---|---|---|
| Claude Personal Finance | From Pro at $20/month (annual billing works out to $17/month, $200 up front), Max $100/month, Ultra $200/month. No separate fee for the feature itself | Beta from Nov 5, 2026, US Pro, Max, and Ultra only. Transactions up to two years; investments and loans down to balances and limits only |
| ChatGPT Finances | From Free at $0 (rolling out to US Free and Go), paid plans separate | May 15 Pro preview → Jun 25 Plus/Pro → Oct 2 Free/Go. More than 12,000 institutions; reads investments and liabilities too |
The gap to read first in this table is not the $20 figure but how far the door has opened. ChatGPT descended over half a year from a preview to the free tier, and reads investments and liabilities as well; Claude starts its beta on three paid plans and reads the investment side only down to balances. What Claude puts forward instead is a seat where money questions carry no ads: no ads, no sponsored recommendations, and no credit cards, loans, or savings accounts of its own to sell, so answers have no product to lean toward — and conversations inside the Personal Finance hub are not used to train models. On the ChatGPT side, connected financial conversations follow the training setting the user chose across all of ChatGPT, so under the same "read-only" label the data defaults point in opposite directions. Which is more trustworthy is a matter of checking, not taste: how settings split inside and outside the hub, and when data is deleted after unlinking — ChatGPT writes deletion of synced data within 30 days, Claude writes an immediate stop to viewing on unlink and automatic unlinking when the account is deleted.
The seat for comparing performance is empty. OpenAI reports 79 points for its default model and 82.5 for its Pro model on an internal benchmark built with more than 50 finance professionals — a company score from the company's own test of its own model. Anthropic publishes no accuracy score for Personal Finance, and no third party has run both products against the same accounts and questions. The performance axis that can be settled today is therefore not a score but scope and defaults: who opens free, how far each reads, and how each keeps conversations out of training.
First, region. The help center's troubleshooting section states that outside the US, or on a VPN that connects from another country, the Personal Finance entry itself does not appear. Whether Korean banks are Plaid-linkable is nowhere in the official materials, so for Korean readers this feature is accurately read as someone else's country's beta, not a household ledger they will soon use. With no Korean launch schedule announced, comparing it with the ledger features of Korean card and bank apps does not hold.
Second, the trap in summing. Reading two years of transactions at once carries seats where transfers between accounts or card payments can be double-counted as spending, and where balances can be a day or two stale. Anthropic's answer is not to guarantee accuracy but to write the checking method into the same documents: ask which transactions sit behind a number, and verify important amounts with your bank. Read-only means money cannot move wrongly — but a plan built on a wrongly summed number remains the user's.
Third, the weight of the connection. Sixteen banks and two years of transactions are, in themselves, a record close to a whole life. Exclusion of hub conversations from training is confirmed, but pull the same numbers into an ordinary conversation outside the hub and your usual privacy settings apply — and the help center also writes that bank links remain in place, after a paid plan ends, until you unlink them yourself. For a feature you turn on because it is convenient, the way to turn it off — per-account unlinking, disconnecting at my.plaid.com, automatic unlinking on account deletion — is part of its real conditions of use, to be checked on the day you turn it on.
The conclusion splits cleanly. Confirmed: the read-only boundary of authority and the Plaid connection; two years of transactions and a 16-bank cap; the reading boundary that investments and loans stop at balances and limits; a paid-only beta from $20/month with no ads and hub conversations out of training as defaults; and the gap that ChatGPT has widened the same capability to free while reading investments and liabilities. Not yet confirmed: how accurate the summing and categorizing are on real accounts, any schedule outside the US or support for Korean banks, and a verified named-user experience. For a reader in the US, the test is small: link one account, ask about one month, match the total against the bank app's own sum to the last cent, and only then open a second account. For a Korean reader, this beta's worth — including that checking method — lies in deciding in advance what to ask when a similar feature arrives domestically.
Puzz, in one line
Claude Personal Finance is a household ledger with a clear design: wide authority to read, no authority to move. But its door is open only on paid US plans, so for Korean readers its first worth is the list of questions to have ready for the day a similar feature arrives, ribbit.