ChatGPT personal finance: what it does, what it misses, and the agent-native alternative.

OpenAI released it as a preview for Pro users in the U.S. on May 15, 2026, and on June 25, 2026 extended it to Plus and Pro subscribers in the U.S. on web, iOS, and Android. You open Finances from the sidebar, or write "@Finances, connect my accounts" in any chat, and ChatGPT walks you through linking accounts with Plaid, which covers more than 12,000 institutions. OpenAI says Intuit support is coming.

Two design choices matter. First, it is read-only: OpenAI states that ChatGPT cannot see full account numbers or make any changes to your accounts. Second, it remembers context you volunteer, such as a savings goal or a loan you owe a relative, as a dedicated kind of memory it calls financial memories, which you can view and delete from the Finances page. Conversations with connected accounts default to GPT-5.5 Thinking.

The strengths are real. Setup takes minutes. The dashboard covers the categories people ask about most: where the money went, what recurs, what is due, how the portfolio moved. Questions like "can I afford this trip" or "which subscriptions should I cut" get answers that cite your actual numbers instead of generic advice. And the whole thing lives inside ChatGPT, so there is no second app to learn.

OpenAI is also candid about the boundary. The launch page says the feature is not a replacement for professional financial advice, and it publishes its own benchmark: the default model scores 79 out of 100 on OpenAI's internal personal-finance evaluation, and the Pro-only model 82.5. That honesty is useful, because it tells you how much of the math to double-check.

It is ChatGPT-only. If you do your real work in Claude, Claude Code, Codex, Cursor, Gemini, or an open-source agent like OpenClaw or Hermes, none of it can see the Finances connection. Your financial context lives in one vendor's chat product, and leaving that product means starting over.

It is a conversation, not an operator. The experience is a dashboard you look at and questions you ask. There is no place to keep an approved budget with line items, no rules that fix a merchant's category every time it appears, no notes an agent leaves for its future self, and nothing that tells you what needs a decision before you think to ask. Financial memories store facts you tell it; they are not a system of record for decisions.

The math is the model's. OpenAI's own benchmark scores the default model 79 out of 100, a composite of expert-graded quality and accuracy on hard personal-finance tasks. That is strong for a language model, and it is still a model producing the number when the question is a monthly total, a pace against a budget, or a projection. Nothing in the experience computes those figures deterministically from the transactions and shows its work.

There is no audit trail. You can see the dashboard, but not a record of which accounts and transactions the assistant read to answer a given question, or why. And on data handling, OpenAI's page says conversations with connected accounts follow the same model training settings you choose across ChatGPT, so unless you have turned training off in Data controls, those conversations are treated like any other. Synced account data is deleted within 30 days of disconnecting, but the financial details already in your conversation history stay until you delete those chats.

The practical difference is who starts the conversation. Candor reviews your records whenever they change and keeps the few situations worth a decision, from a subscription whose price just rose to cash sitting idle, budget drift, or an employer match left on the table, each with the figures behind it. If nothing qualifies, the list is empty. Your agent reads that list when you ask, or on a schedule if it supports scheduled tasks, say every Monday morning, and brings you the decision. Nobody has to remember to ask.

The math difference is architectural. Candor computes totals, budget pace, and variance server-side from the transactions and returns them with the records behind them, so the agent reports a figure instead of predicting one. The trust difference is the same shape: Candor holds no money-movement tools at all, is read-only toward your institutions by design, logs every read with a reason, never sells or trains on your data, and lets you export your data and delete it. Your agent still makes the judgment calls and acts under your authority; it just does so from a workspace that remembers.

Plenty of people will keep ChatGPT Finances for quick questions and give Candor to the agent that does their scheduled work. If ChatGPT is that agent, the setup page below covers the routes, none of which needs a terminal.

The difference is what happens to a question after you ask it. Ask ChatGPT Finances which of your subscriptions went up and you get an answer for that chat. Give the same question to an agent with Candor and it stays on watch: price history on every recurring charge, the increase on the list of what needs a decision the day it posts, a note recording whether you kept or cancelled it, and a check next cycle that the charge actually stopped. The same shape holds for the questions people ask most. Is any of my cash sitting somewhere that earns nothing? Which card is closest to its limit, and what brings it under 30%? Project my balances 60 days out and show me the week money gets tight. Did all the income I expected actually arrive, at the right amounts? What am I paying in fund expense ratios across my investments? If I had spare money each month, which debt first? In Candor each of those starts from figures computed from the records, with the transactions attached, and the answer, the decision, and the follow-up survive the conversation.

The honest tie-breaker is the second month. ChatGPT Finances is simpler if you expect to open a dashboard and ask a few things. Candor is built for the version where the work happens without you: a scheduled check-in, where your agent supports one, that starts with what needs a decision, a duplicate charge caught and the refund followed until it landed (the case study below recovered $1,149.69), a stress test of how long your cash lasts if income stopped, and a history your next agent inherits the day you switch.

Candor is included because this is Candor's site. Where ChatGPT Finances is the better fit, the page says so. If OpenAI changes the feature, this page will be updated with the date.