ChatGPT Is Already Advising You on What to Do With Your Money. Can You Trust It?

A solo ToTheMoon episode with Alexander Volchek on whether artificial intelligence can be trusted with financial advice and with everything to do with handling money. A year or two ago such a question would have sounded like hype: people were uploading reports, accounts, contracts and tax data into AI, and the channel had plenty of cases where the systems worked better than financiers, accountants and lawyers. But now, at the end of September 2026, there is a fresh study by the company Saturn, "Artificial Authority. Should You Trust Artificial Intelligence With Financial Advice?", on fresh models — Grok 4.5, ChatGPT 5.6, Gemini 3.1 Pro, Claude Opus 5 — and its error rate made Alexander record a separate episode.

He himself advises asking models financial questions every day — about accounts, credit ratings, investments, taxes, bookkeeping — and recently told how he analysed all his accounts and reporting through Codex, so that systems such as QuickBooks may become unnecessary. And right before the recording ChatGPT itself suggested he connect his credit rating via Experian: this is GPT Finance, a financial module, or plugin, that connects accounts at more than 12,000 financial institutions via Plaid, and not only in the US. You can ask about money through such plugins and connectors, or directly — in the chat or in Codex.

Saturn took 18 models — ChatGPT, Claude, Gemini, Grok, Copilot — and asked each of them 121 financial questions five times: more than 10,000 answers, but these are repeated runs of a small set of questions about debts, loans, mortgages, pension savings, taxes, savings and student loans. According to Alexander, it was the US market that was studied, where the systems have the most integrations, but models can know well-documented countries very well. An answer failed for a factual error or for missing an essential point or a necessary warning.

The result: 57% errors on average, 63% for free models, 49% for paid ones, 88% on the complex questions and 93% for free models on the hardest ones. At the same time the paid versions were cheap ones: no 5.6 Sol Pro, no 6.0 Astra Pro, no Fable 5.1, no Extra High or Max modes; a Pro version, by Alexander's estimate, would be wrong some 20–30% of the time. To those who begrudge $20 for a subscription he points out that today this buys a few cups of coffee or a kilo of meat, and urges them to understand the risks of the tools they use.

The errors are concrete. Claude Haiku 4.5's advice on pension taxes could, by the authors' estimate, have led to an additional tax charge of £17,500; a model advised paying off the highest-rate debt first without taking rent, taxes and arrears into account; Claude invented a student-loan rule for people who move abroad; Gemini assured a borrower that a mortgage payment holiday would not affect their credit score — in the US that can cost several years of a record. Sam Altman said a year ago that a model should not be trusted without limits, and ChatGPT itself says it can make mistakes, yet people act on autopilot.

By model: Claude Haiku 4.5 — 82%, Gemini 3.1 Pro — 73% despite all of Google's integrations and knowledge base, Grok 4.5 — 60%, ChatGPT 5.6 Luna — 58%, and the report does not say which mode Luna ran in; Alexander himself does not use such models. The best result went to Claude Opus 5 in reasoning mode, 39%, but even that is a big question for a serious modern model. A separate category of errors: the models ignored upcoming changes, tax changes for example.

Connecting data does not solve everything: GPT Finance breaks down spending, builds budgets and keeps a memory — memories — but it does not know about your illness, a unique exception, an inheritance or a sudden change in income, and accountants, banks and lawyers give differing information themselves. Alexander does not like plugins because he does not fully understand which rules are set up inside them; Claude has connectors and a new system for financial advisers, but connecting data does not mean granting access to payments — the system can be given analysis alone. Third-party finance apps run on cheap versions — he himself got a report for one of his companies from a system that Codex wrote to bypass QuickBooks.

The upshot: the figure struck Alexander, and for the coming months he may be even more careful with finances through AI models, although he still considers these systems smarter than any person at working with data. What matters is which questions you ask them, how you phrase them and whether you double-check the answer.

The value of the episode is that the argument about trusting AI with money is turned into figures — 57% errors on average, 88% on the complex questions, 39% even for the best model — and into concrete scenarios where an error costs money: from an additional tax charge to several years of a record in a credit history. The host does not give up on AI in finance — he gets the reporting for one of his companies from a system Codex wrote and considers the models smarter than any person at working with data — but insists on strong models, double-checking and questions from which a model learns about you what is not in the connected accounts. The practical conclusion: connected financial data makes an answer more convenient, but it does not make it correct.