September 2026 is the month the chatbots officially came for your wallet. On September 21, OpenAI switched on credit-score tracking inside ChatGPT — connect an Experian report and your score lives in your chat history, with monthly refreshes and alerts. Two weeks earlier, the AP examined OpenAI’s rolling Finances feature, which links your bank accounts, cards and brokerages through Plaid, as anti-money-laundering experts warned what happens when phishing meets a chatbot that knows your net worth. In between, code watchers caught Anthropic quietly testing “Claude Money” — its own bank-linking feature — inside the Claude iOS app. And on September 19, the Financial Times published the study that should sit next to all of this news like a warning label: after running 18 AI models through 121 money questions more than 10,000 times, researchers found chatbots got financial answers wrong 57% of the time — and wrong 88% of the time on hard, tax-heavy questions.
If you freelance, run a one-person business, or earn in dollars while living abroad, this collision matters more to you than to anyone. You have no HR department, no benefits desk, no colleague in accounting to sanity-check you. ChatGPT has become, by default, the closest thing most solo earners have to a finance department — OpenAI itself says more than 200 million people a month come to ChatGPT for budgeting, investing questions and planning. Now the tools are getting deeper access to your actual money at the exact moment the evidence says their advice is unreliable.
What you will get in this guide:
- Everything that actually shipped: ChatGPT Finances, Experian credit tracking, the Privacy Center, and Claude Money
- The Saturn “Artificial Authority” study in full — the numbers, the worst errors, and why free users get the worst advice
- The privacy and security blast radius of linking accounts to a chatbot
- Nine concrete moves for using AI finance tools as a freelancer or digital nomad without getting burned
What ChatGPT Finances Actually Is — and How We Got Here
ChatGPT Finances is a dedicated area inside ChatGPT where you can securely connect financial accounts, see a dashboard of your money, and ask questions grounded in your real balances and transactions. It launched as a Pro-only preview in May, went to all paying U.S. users in June, and has been expanding quietly ever since. The full timeline:
| Date (2026) | What shipped |
|---|---|
| May 15 | Finances preview for U.S. ChatGPT Pro subscribers (web + iOS) — first time a general chatbot could see live bank data at scale |
| June 25 | Rollout to all U.S. Plus and Pro users on web, iOS and Android — the feature reaches the ~$20/month mass market |
| September 14 | AP-syndicated coverage of the gradual rollout; ACAMS anti-money-laundering director warns about account-takeover risk |
| September 16 | TestingCatalog spots “Claude Money” — a bank-linking feature in testing inside the Claude iOS app |
| September 19–21 | FT reports the Saturn study: AI models wrong on 57% of money questions — the first large-scale accuracy audit of consumer AI financial advice |
| September 21 | Experian credit-report + VantageScore 3.0 tracking arrives in ChatGPT Finances; OpenAI ships a sitewide Privacy Center — credit-bureau data now lives inside a chat history |
Here is what the connected experience does, per OpenAI’s own documentation. Accounts link through Plaid (Intuit support is “coming soon”), covering more than 12,000 U.S. financial institutions with read-only access. Once synced, you get a dashboard spanning portfolio performance, spending, subscriptions and upcoming payments. Type “@Finances, help me save more over the next few months” and the model — GPT-5.5 Thinking by default — reasons over your actual transactions and produces a category-by-category plan. A Financial memories layer stores context you share (saving for a car, a loan from your parents) to inform future answers.
Just as important is what ChatGPT cannot do. It cannot see full account numbers, move money, pay bills, trade securities, change account settings, or file taxes. OpenAI’s documentation states the service does not act as a financial adviser, and the announcement page carries the line: “ChatGPT can help you stay informed and feel more confident managing your finances, but it is not a replacement for professional financial advice.” The Intuit partnership sketches where this is heading — from a card recommendation to a submitted application, from a tax question to a live tax professional — but today the product is strictly read, summarize, suggest. OpenAI has also published its own accuracy homework: an internal benchmark graded by more than 50 finance professionals, on which GPT-5.5 Thinking scores 79 out of 100 and GPT-5.5 Pro scores 82.5. Hold those numbers in one hand; in a moment we will put a very different set in the other.
Credit-Score Tracking: What OpenAI Switched On September 21
The newest layer deserves a slow read. Eligible U.S. Plus and Pro subscribers can now connect an Experian credit report and see an Experian-based VantageScore 3.0 inside Finances. OpenAI’s announcement on X was breezy: “You can now track your credit score in ChatGPT, and stay informed when things change.” The mechanics are less breezy:
| Item | Detail | What it means for you |
|---|---|---|
| Score shown | Experian-based VantageScore 3.0 | Not necessarily the score a lender uses — mortgages typically pull FICO, often from a different bureau |
| Inquiry type | Soft inquiry | Checking your score in ChatGPT does not lower it |
| Updates | Monthly report + score refresh | Recent activity may not show for weeks; it is a monitor, not a live feed |
| Alerts | New inquiry, new account, new address | A fraud tripwire — treat each alert as a prompt to inspect the report, not proof of anything |
| Verification | Experian identity check required | Only complete it inside the official flow, on a trusted device, logged into the right account |
| Bank links | Experian connection is separate from Plaid | You can enable credit monitoring without linking a single bank account |
| Availability | U.S., Plus ($20/mo) and Pro ($200/mo) only | Free users and everyone outside the U.S. get nothing — for now |
Two fine-print items matter enormously. First, deletion is asymmetric: disconnecting Finances deletes synced account data within 30 days, but information already woven into your conversation history is not automatically removed — balances and spending patterns can persist in old chats you have forgotten about. Second, OpenAI shipped a Privacy Center the same day: a single dashboard for chat privacy, memory, personalization, data use, connected apps and account security, open to all tiers including Free. The timing is not subtle — OpenAI is tidying the privacy house in the same week it starts pulling credit-bureau data through the front door.
For freelancers there is a specific trap in the VantageScore detail. When you apply for a mortgage as a self-employed borrower, lenders scrutinize the exact FICO score and bureau file they pull — a point we walk through in our guide to getting a mortgage when you are self-employed. A ChatGPT score that looks healthy can coexist with a lender-facing score that does not. Treat the chatbot view as a trend line, never as the number you bet a loan application on.
Claude Money: Anthropic Is Two Steps Behind — and Moving Fast
On September 16, BleepingComputer reported what TestingCatalog found buried in the Claude iOS app: a new Money section alongside Chats, Code, Artifacts, Dispatch and Cowork, with the tagline “Understand your money with Claude. Link your bank accounts and ask Claude about spending, plans, and more” and a Get started button that is not yet lighting up for most users. Anthropic has announced nothing publicly; supported banks, the account aggregator, and country availability are all unknown. BleepingComputer’s read is that Europe may never see it at all, because GDPR-style privacy law makes chatbot-meets-bank-data a much harder sell there.
The competitive picture as of this week:
| Assistant | Maker | Bank account links | Built-in credit tracking | Status |
|---|---|---|---|---|
| ChatGPT (Finances) | OpenAI | Yes — Plaid, 12,000+ institutions, read-only | Yes — Experian + VantageScore 3.0 (U.S., paid tiers) | Live and expanding |
| Claude (Money) | Anthropic | In testing on iOS | Not announced | Spotted in app code, not rolled out |
| Gemini | No consumer bank-link feature | No | Finance features live in enterprise products only | |
| Copilot | Microsoft | No | No | Productivity focus; finance via enterprise Copilot |
OpenAI currently has the only live consumer chatbot integration with a credit bureau — a first-mover position Google and Microsoft have conspicuously not contested. If the rollout stays incident-free, expect the bureaus to cut similar deals with everyone else within a year or two.
The 57% Problem: What the Saturn Study Actually Found
Now the other hand. Saturn, a UK-based AI firm serving more than 750 financial-advice firms and 6,500 advisers, published the “Artificial Authority” report — the most thorough public benchmark yet of consumer chatbots on money questions. The method: 18 AI models (free and paid versions of ChatGPT, Gemini, Claude and Copilot), 121 questions spanning debt, student loans, mortgages, pensions, inheritance tax and retirement planning, each run five times — more than 10,000 assessed responses. Grading was strict all-pass: every required figure, rule, warning and deadline had to be present.
| Measure | Result |
|---|---|
| Average accuracy, 17 current models | 43% — wrong or materially incomplete 57% of the time |
| Easy questions (financial literacy, no calculation) | 54% accuracy |
| Hard questions (multi-part tax scenarios, precise figures) | 12% accuracy — 88% wrong |
| Failure rate: free vs paid models | 63% vs 49% |
| Worst free model | Claude Haiku 4.5 — wrong or incomplete 82% of the time |
| Best free model | ChatGPT-5.6 Luna (max) — still wrong or incomplete 56% of the time |
| Best model overall | Claude Opus 5 (reasoning), paid — 61% pass rate; 67% wrong on hard questions |
The error taxonomy is worse than the headline. Of all failures, 37.1% were incomplete beyond a single missing fact, and 18.4% left out a required figure, limit or deadline. But 6.1% involved fabricated rules — confident inventions of law that does not exist — and 2.3% served up out-of-date regulatory guidance. Saturn stresses the detail that makes this dangerous rather than merely annoying: wrong answers arrive in the same fluent, authoritative tone as right ones. There is no hedging tell that flags a hallucination.
The report documents examples with real dollar consequences. One model invented a rule that a graduate could halt federal student-loan payments by moving abroad — in reality, doing so without an approved deferment or income-driven plan leads to delinquency, default, collections, and potentially higher payments. A Gemini model told a borrower that a mortgage payment holiday would not affect their credit score — wrong, and the kind of error that costs a five-figure rate on the next refinance. Saturn also flags the equity problem: consumers least able to pay for advice are most likely to rely on free AI — and free models failed 63% of the time.
Meanwhile the demand curve keeps climbing. Intuit Credit Karma survey data cited by Tom’s Guide found 66% of Americans who have used generative AI have used it for financial advice, rising to 82% of Gen Z and millennials, with finance the second most common GenAI use case (41%) behind health and wellness. EY’s 2026 research found roughly half of respondents already use AI for saving or investing. Millions are asking; the models are missing more than half the time.
Wait — didn’t OpenAI say 79/100?
Yes, and both numbers can be true. OpenAI’s benchmark is a weighted composite of expert-graded quality and accuracy on tasks its models were tuned for. Saturn graded strict factual completeness, where one missing deadline fails the whole answer. Different instruments — but notice what they agree on: even the friendliest measurement leaves a fifth of the performance on the table, and the strictest says hard questions fail almost nine times out of ten. Neither number describes a tool you should let make an irreversible financial decision unattended. OpenAI, to its credit, says the same about its own product.
Why This Hits Freelancers and Digital Nomads Harder
An employee who asks ChatGPT a bad 401(k) question has a plan administrator, an HR portal and a matching-contribution default protecting them from their own curiosity. A freelancer has none of that — and the questions freelancers ask are structurally harder, landing squarely in the 12%-accuracy zone: Should I make this quarter’s estimated payment on annualized income or last year’s safe harbor? Does moving tax residency to Portugal trigger the foreign earned income exclusion or the Portuguese regime? Is this income U.S.-sourced if the client is American but I was in Bali when I did the work? These are multi-jurisdiction, interacting-rules, precise-deadline questions — precisely what Saturn found models fail 88% of the time. Our guides to quarterly estimated taxes and the foreign earned income exclusion exist because getting these wrong is expensive; an AI that fabricates rules in a confident tone is a penalty generator.
Then there is the security blast radius. Craig Timm, senior director of anti-money-laundering at ACAMS, told the AP the quiet part out loud: if someone takes over your ChatGPT account — via a data leak, a phishing email, a reused password — they inherit a curated financial dossier on you: balances, spending habits, subscriptions, credit report. His specific worry is precision social engineering: an attacker who can see your transactions can pose as your doctor’s office or your bank, reference your real activity, and walk you into handing over full account access. His take on consent fatigue: “The privacy notice comes up… ten pages in really small print, and we just say, whatever, this is a battle I’m not going to fight.” For a nomad working from hotel Wi-Fi, that attack surface travels with you.
There is also a durability problem unique to location-independent earners. Finances and credit tracking are U.S.-only features tied to U.S. institutions and bureaus. Your foreign bank accounts will not link through Plaid; spend a year in Europe and your “AI finance department” quietly becomes a plain chatbot again — and Claude Money may never launch in the EU at all. Building your financial operating system on a feature that can geofence you out mid-year is a planning error. The stack that never locks you out — multi-currency accounts, a real accounting platform, and a high-yield savings home base — stays boring and stays yours.
The 9-Move Playbook: Using AI Finance Tools Without Getting Burned
Move 1: Give AI the job it can do — bookkeeper, not adviser
Frame every interaction around the role the evidence supports. AI is genuinely good at aggregation and pattern-spotting: categorizing three months of messy transactions, surfacing the $14.99 subscription you forgot, showing that client A pays in 9 days on average while client B pays in 47. It is not good at binding rules with deadlines and penalties. The split worth memorizing: AI organizes my data; licensed humans and official sources make my decisions.
Move 2: Connect the minimum scope — and know you can connect credit only
September 21’s design has an underappreciated gift: the Experian credit connection is separate from Plaid. You can get score tracking and change alerts without linking a single bank account. If you do link accounts, start with one or two (checking + primary card), skip the brokerage, and skip accounts holding client escrow money. Every account you do not connect is data that cannot leak or be phished out of an account takeover.
Move 3: Fortify the front door before you let data in
Before connecting anything: give your ChatGPT account a unique password from a password manager, enable multi-factor authentication (OpenAI explicitly recommends it for Finances users), and spend ten minutes in the new Privacy Center. Check two things — whether your conversations are used for model training (turn it off if you want), and the full list of connected apps with account access. The AP’s takeover scenario assumes a weak door; make yours boring to break into.
Move 4: Keep the toxic three out of every prompt
Never type full account numbers, government ID numbers (Social Security, passport), or client-identifying contract details into a chat. The models do not need them — connected data arrives through Plaid, and questions can be asked generically (“a client owes 18,000 across two invoices”). For one-off sensitive questions, use a temporary chat: by design it cannot access connected accounts and leaves nothing in history. Disconnecting accounts does not scrub old conversations — the cheapest privacy win is never putting sensitive text into the history at all.
Move 5: Apply the hard-question rule — nothing with a deadline, dollar figure or tax citation leaves the chatbot unverified
This is where 88% wrong becomes 0% wrong: treat every AI answer involving a tax rule, penalty, deadline or threshold as a draft. Ask the model to cite its basis, then confirm on IRS.gov, SSA.gov, your servicer’s site, or the official page for your host country. Fabricated rules were 6.1% of all failures — plausible inventions that read like law. Verification costs ninety seconds; the unverified version costs the underpayment penalty we broke down in our estimated-taxes guide, or a defaulted loan from a hallucinated rule.
Move 6: Use the two-model cross-check for anything over $1,000
For consequential decisions — taking a client’s equity instead of cash, choosing between solo 401(k) and SEP-IRA, timing a large equipment purchase — run the same question through a second model and compare. Divergence is information: where two frontier models disagree, a human is required. The best model in the Saturn study still failed almost four times in ten; a second opinion plus a fee-only CFP or CPA is how you turn 61% confidence into a decision you can defend. Our retirement planning guide shows trade-offs a model will gloss over.
Move 7: Put the dashboard to work on freelancer-specific patterns
Irregular income is where connected-data AI genuinely shines: spotting patterns across many transactions is exactly what lumpy cash flow needs. Good uses this month: reconcile your emergency fund runway against real trailing-three-month expenses instead of a guess; find the subscription stack quietly costing a freelancer’s worth of groceries; identify which clients’ payment delays are financing your overdraft fees; and draft the rate-increase email you have been avoiding, with your own payment-history data as the argument. Bad uses: which stock to buy, whether to skip an estimated payment, what a tax rule means for your residency situation.
Move 8: Run the 20-minute monthly reconciliation ritual
Once a month: pull the AI dashboard’s category totals, export the same month from your bank as CSV, and make the two agree. Discrepancies teach you what the AI mis-categorizes (often transfers, refunds, foreign-currency conversions) and what your own mental bookkeeping missed. Then check the Experian alerts ChatGPT surfaces — a new inquiry you did not initiate or an unfamiliar address is a signal to pull the full report at annualcreditreport.com and, if it looks real, freeze your credit with all three bureaus. Alerts are prompts to verify, never proof. If you are building credit from abroad, follow the method in our credit score abroad guide — a VantageScore in a chat window is not the same asset as a lender-ready FICO file.
Move 9: Build the nomad layer — assume the feature disappears when you move
Treat ChatGPT Finances as a U.S.-resident luxury, not infrastructure. Your primary stack must work from any country without a chatbot: multi-currency accounts that hold and convert properly, an accounting platform that is the system of record, a tax plan maintained by a human who knows your host country, and quarterly statement backups exported from the AI dashboard. Then the day a geofence, a policy change or a privacy incident turns the feature off — or the day Claude Money ships and you switch — your financial memory walks out with you.
Frequently Asked Questions
Is ChatGPT Finances safe to use?
Reasonably, with scope discipline. Access is read-only through Plaid, ChatGPT cannot see full account numbers or move money, synced data is deleted within 30 days of disconnecting, and OpenAI says it does not train models on your connected personal data — conversations follow whatever training setting you choose, which you can turn off. The residual risks are manageable: account takeover (MFA + unique password), sensitive text persisting in chat history (do not type it), and wrong answers (Moves 5 and 6).
Does checking my credit score in ChatGPT lower it?
No. The Experian connection uses a soft inquiry, which does not affect your score. Monthly alerts (new inquiry, account, or address) help you catch problems early.
Why does ChatGPT show a different score than my lender?
ChatGPT displays an Experian-based VantageScore 3.0. Lenders commonly use FICO scores, other VantageScore versions, or different bureaus — different model plus different data produces different numbers. For a mortgage or major application, ask the lender which score and bureau they use.
Is Claude Money available yet?
Not publicly. It was spotted in testing inside the Claude iOS app in mid-September 2026, but Anthropic has announced no launch date, no bank list, and no countries — and European availability is uncertain given privacy law.
Can ChatGPT replace a financial adviser?
No — and OpenAI says so on the product page. The independent evidence is harsher than the disclaimer: 43% average accuracy, 12% on hard multi-part tax questions, fabricated rules in 6.1% of failures, and the best model tested still failing almost four in ten. An adviser is a fiduciary with malpractice exposure; a chatbot is neither. Use AI to prepare better questions for the human, not to replace the human.
How much does this cost?
Finances and credit tracking require a paid U.S. subscription: ChatGPT Plus at $20/month or Pro at $200/month. There is no separate fee for the Experian connection, and the Privacy Center is available to everyone, including free users.
The Bottom Line
September 2026 drew the line between two eras. Before: chatbots guessed at money questions from general knowledge, and you could sort of tell. After: they can see your actual balances, your actual credit report, your actual subscriptions — and they still guess, with the same fluent confidence, 57% of the time on average and 88% of the time on the hard questions freelancers live inside. Deep data access did not make the advice good; it made the advice personalized, which is far more persuasive and equally often wrong.
The winners of this transition will be solo earners who treat the new tools with professional detachment: connect the minimum, lock the door, keep the toxic three out of prompts, verify every deadline and dollar figure against primary sources, cross-check big decisions with a second model and a licensed human, and keep the boring non-AI stack as the system of record. Do that and the chatbot becomes what it should have been all along: the fastest bookkeeper and pattern-spotter a one-person business has ever had. Skip it, and you are one confident hallucination away from a missed estimated payment, a damaged credit file, or a phishing attacker who knows your finances better than your bank does. The chatbot belongs on top of your money stack — never underneath it.
Building your money stack for location independence? Continue with our guides to the best investment apps for expats, paying off credit-card debt on irregular income, and AI-proofing your freelance income — the other half of the same story.
Sources: OpenAI, “A new personal finance experience in ChatGPT” (openai.com, updated June 25, 2026); AP via DC News Now, “OpenAI rolls out personal finance tool, as experts warn of privacy concerns” (September 14, 2026); BleepingComputer, “Anthropic wants Claude to analyze your bank account and financial data” (September 16, 2026); InvestmentNews, “AI chatbots give wrong financial answers most of the time, study finds” (September 21, 2026), on Saturn’s “Artificial Authority” benchmark (saturnos.com); Tom’s Guide, “AI could be costing you money” (September 21, 2026), with Intuit Credit Karma and EY survey data; TechJournal, “ChatGPT Credit Score Tracking Comes to US Subscribers” (September 21, 2026); shattered.io, “ChatGPT Credit Score Tracking & Privacy Center” (September 22, 2026); Financial Times (September 19, 2026). Figures are as reported by the cited sources; data current as of September 23, 2026.