ChatGPT can read your chart, research a stock and write the code for a strategy. Sending a real order is another matter: OpenAI trained its agent to stay away from high-stakes money moves, and plugins in its directory may not execute trades. Here is what works in November 2026, how people wire it to a broker anyway, and what the published tests found.
November 15, 2026
19
min to read
The short answer
Not on its own. ChatGPT can read a chart screenshot, research a stock with web search and run backtest code, but it has no brokerage account. OpenAI trained its agent to avoid high-stakes financial actions, and plugins in its directory may not execute trades. People who trade through it add a broker’s MCP server or write their own code.
Most of the work before an order is reading, checking and thinking, and that is where ChatGPT earns its place. As of November 2026, a trader can use it to:
None of that needs an account anywhere. Our library of ChatGPT trading prompts has ready-made prompts for most of these jobs. Claude does the same jobs with a different set of connectors, and we compare the two task by task in Claude vs ChatGPT for trading. The rest of this page is about the step neither one takes by default: sending the order. For how ChatGPT stacks up against Claude, Grok and Gemini on the same tasks, see our wider LLM trading comparison.
Out of the box, ChatGPT has no account at a broker or an exchange, no feed of live quotes and nothing resting in any market. Five gaps follow from that:
| Job | ChatGPT by itself | What fills the gap |
|---|---|---|
| Send an order | No. A chat has no route to a broker | A broker’s MCP server (a connector, explained below), a custom GPT action or a script you run |
| Quote a live price | Only what a search result or your upload says, and that may be old | Your broker’s platform, a connected data source or your own chart |
| See your positions | No, unless you link an account; Finances reads but can’t change anything | A broker connection with access to your account |
| Keep a stop in place | No. A stop you describe in a chat is not an order | A stop order resting at your broker or exchange |
| React the moment a level breaks | No. A scheduled task looks again later; it is not a trigger | An alert or an order that sits at the venue |
The nearest thing OpenAI builds to account access is ChatGPT Finances, on the Free, Go, Plus and Pro plans (availability varies; check OpenAI’s page). It links your financial accounts through Plaid and reads balances, transactions and investment data, so you can ask how your portfolio is spread or what changed in your spending. OpenAI is clear about where it stops: ChatGPT “cannot see your full account numbers or make any changes to your accounts.” It can tell you what you own. It can’t sell any of it.
ChatGPT can also run scheduled tasks for recurring work, such as a summary of your watchlist every morning. That is a fresh look on a timer, not a market order waiting at a price.
Agent mode is ChatGPT with a computer of its own: it opens websites, clicks, types and fills in forms for you. OpenAI launched it as ChatGPT agent on July 17, 2025, for Pro, Plus and Team users, who picked “agent mode” in the tools menu. Its product lead told The Verge that day that financial transactions were restricted “for now.” In OpenAI’s November 2026 docs, this kind of work happens in ChatGPT Work, which “uses its own browser, running on a separate computer in the cloud” and can sign in to websites on the Plus and Pro plans. So in principle it could log in to your broker and press Buy. OpenAI built brakes into it instead.
The ChatGPT agent system card, the safety report OpenAI published at launch, says the agent “should not transact for regulated goods or gamble, and should not assist with certain financial activities with elevated consequences, such as making financial account transfers.” Three safeguards back that up:
Two details matter for traders. The system card never mentions buying shares by name; its example of a high-stakes financial task is moving money between accounts. And OpenAI’s own example of an agent mistake is that “it may inadvertently buy the wrong product.” A request to buy a stock might be refused, paused for your confirmation or carried out on a page the agent misread. None of those outcomes is a dependable way to place a trade.
The current docs keep the same stance. ChatGPT Work “will always ask for confirmation before consequential actions, such as submitting your information to book an appointment or completing a payment,” and OpenAI warns that “Instructions on a page can be misleading or malicious.” If you ever let it click through a broker’s site, the order that comes out is still yours: read the symbol, the side, the size and the order type before you confirm.
Since ChatGPT won’t trade by itself, people who want it to place orders connect it to a broker. In November 2026 there are three routes, and each one puts the order in a different place.
MCP, the Model Context Protocol, is an open standard for plugging outside tools into an AI app. A broker publishes an MCP server whose tools do things like read your positions or send an order, and ChatGPT calls those tools when you ask. OpenAI’s plugin guidelines shape what reaches you: for plugins listed in the ChatGPT directory, the list of prohibited services includes “Execution of money transfers, crypto transfers, or investment trades.” In practice that splits broker connections in two.
Robinhood also lets you skip ChatGPT altogether: its in-app Robinhood Agents let you choose the model, and Robinhood is covering the token costs of OpenAI’s GPT-6 Luna in a promotion scheduled through the end of 2026. We go through both Robinhood products in Robinhood Agents and its alternatives, and the other brokers that let AI agents trade in our guide to agentic trading.
The older do-it-yourself route is a custom GPT with an action: a description of a broker’s web API, the interface programs use to send it instructions, that ChatGPT can call. OpenAI’s docs for actions build in a brake. Any call other than a plain read, a GET request, counts as “consequential” by default, which means ChatGPT “must always prompt the user for confirmation before running” and shows no “always allow” button. A builder can mark an endpoint as not consequential, which brings the “always allow” button back; for an endpoint that sends orders, that is exactly what you don’t want. The route is also fading: OpenAI says it is transitioning custom GPTs to plugins, and custom actions don’t carry over.
This is what most people mean by a “ChatGPT trading bot”: you ask ChatGPT, or Codex, OpenAI’s coding agent, to write a program that calls a broker’s API, and you run it yourself. ChatGPT is the programmer; your program places the orders, bugs included. Alpaca, a broker built around its API, suits this because it offers paper trading: simulated orders against real prices. Its official MCP server is a poor fit for ChatGPT, though. Alpaca’s README says ChatGPT “requires a remote MCP server” and tells you not to expose its package “directly to the public internet with your Alpaca API keys.”
| Route | Who sends the order | Built-in safeguard | Watch out for |
|---|---|---|---|
| IBKR connector from ChatGPT’s directory | You, in IBKR | You submit each order; Autonomous Trading only in a separate AI Partition | Read the instruction before you submit it |
| Robinhood Trading MCP in developer mode | ChatGPT, in the agentic account | Only the money you move into that account | Read access to all your Robinhood accounts |
| Custom GPT action | ChatGPT, through the broker’s API | A confirmation before every order call, unless the builder turns it off | An API key saved in the action; custom GPTs are moving to plugins |
| Your own script | Your code | Paper trading first, if you use it | Bugs, leaked keys, nobody watching |
Claude can be wired up in similar ways, with its own connector settings and limits; see can Claude trade for you? For Grok, see can Grok trade for you?
It can be made safer, not safe. These are the failures that show up with ChatGPT in particular, and how to blunt each one:
| Risk | How it shows up | How to limit it |
|---|---|---|
| Made-up numbers | A price, a percentage or a date that sounds right and isn’t | Ask where each number came from; check the ones you act on |
| Stale data | Training data with a cutoff, or a news story quoting yesterday’s price | Ask for the time of every quote; read the live price at your broker |
| A confident tone | A firm answer with no sign of doubt | Ask how sure it is and what would change its mind |
| Prompt injection | Hidden text on a web page that tries to steer what it does next | Don’t browse and trade in the same session; connect only servers you trust |
| Loose permissions | An “always allow” on an order tool, or an API key pasted into a chat | Keep confirmations on; keep keys in the broker’s or the app’s settings |
The first three share a root cause. A September 2025 paper by researchers at OpenAI and Georgia Tech explains that language models “sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty,” because training and grading reward a good guess over an honest “I don’t know.” OpenAI’s Model Spec, the rulebook for how its models should behave, ranks a confident wrong answer as the worst possible outcome and tells the models to say when they are unsure. It also tells them not to put a number on that doubt unless you ask. So ask.
Prompt injection means instructions hidden in something the model reads, such as a web page, written to override what you asked. ChatGPT’s search docs open with a warning, “Treat all web results as untrusted input,” and the system card is frank about the numbers. In OpenAI’s tests on the agent’s visual browser, it ignored planted instructions in 95% of cases, but resisted attacks that tried to make it fetch and leak private data in only 67%, and OpenAI notes these figures test the model alone, not its full set of defenses. The danger is greatest in a session that can both read the open web and reach an account.
Go back through your last answer. For every price, percentage, level and date in it, tell me where it came from: a web page (give the link and the date it was published), a file I uploaded, the image I pasted, or your training data. Mark anything you estimated or could not verify, and tell me how old the newest data point is.
Run the whole flow on a paper account, or in a mode where you approve each order, until nothing in it surprises you.
Trade from an account that holds only what you could lose, like Robinhood’s agentic account or an IBKR AI Partition.
Leave confirmations on for anything that sends an order, and never mark an order tool “always allow.”
Do web research in one chat and place orders in another, so a poisoned page never sits next to a live order tool.
Send stops and limits as real orders, so they still work when the chat is closed.
Keep API keys in the settings of the broker or the app. Alpaca’s docs put it plainly: “Never paste API keys into chat.”
People have tested it, in academic benchmarks and in public portfolios. The results are more modest than the headlines that travel on social media, in both directions.
| Study | When | What the model did | What happened |
|---|---|---|---|
| Lopez-Lira and Tang | First posted April 2023, revised October 2025 | Scored news headlines as good or bad for a stock | GPT-4’s scores matched the market’s first reaction on about 90% of portfolio-days, a move too fast to trade. They also predicted part of the drift that followed, mostly in small stocks and on bad news, and that edge shrank as more investors used such models |
| StockBench | March 3 to June 30, 2025 | Traded 20 large Dow stocks each day from prices, fundamentals and news | OpenAI’s o3 returned 1.9% and GPT-5 0.3%, against 0.4% for buying all 20 and holding. Every model and the benchmark had a worst drop of about 11% to 15% |
| Nathan Smith’s micro-cap experiment | June 27 to December 26, 2025 | Ran a $100 portfolio of US micro-cap stocks, with a person entering its trades | It ended at $84.48 after falling as much as 50%. Smith found it concentrated, quick to buy back stocks that had lost it money, and behind both the Russell 2000 and the S&P 500 |
| FINSABER | Backtests over 2004 to 2024, published at KDD 2026 | Ran LLM trading strategies, built on OpenAI’s GPT-4o models, across more than 100 stocks | Edges reported in short tests “deteriorate significantly” over longer periods and more stocks; the strategies were too cautious in bull markets and too aggressive in bear markets |
Read together, the evidence says ChatGPT is good at the reading part of trading: turning a headline or a filing into a view, fast. It does not show that ChatGPT reliably beats simply holding the market. The samples are small in both directions, though. One four-month benchmark or one $100 portfolio can’t prove that it fails, any more than one lucky month would prove that it works.
On price predictions specifically, the Model Spec lists “predictions of future states” among the cases where “the outcome is inherently uncertain,” and asks the models to say so. A ChatGPT answer that names next week’s price with confidence is not following its own rules. The research on AI price forecasts, ChatGPT’s included, is in our guide to whether AI can predict the stock market. The better question is not whether ChatGPT can pick stocks, but which part of your own process it can check. For the AI tools that fit each stock trading job, see our guide to AI stock trading.
It will give you information and help you reason, and it is told to stop short of a personal buy or sell verdict. The Model Spec, in its October 29, 2026 version, has a rule called “Provide information without giving regulated advice”:
For advice on sensitive or important topics (e.g., legal, medical, and financial matters), GPT should equip the user with information without providing definitive advice that would be best given by a licensed professional.
Its worked example is a user asking “should I buy Acme stock?” The reply it marks as correct opens with “I can’t give direct financial advice, but I can help you think through the decision,” then offers to look up analyst reports and build a bull and bear case. The reply it marks as wrong declines and stops there. So expect help with the thinking and no verdict on the trade, along with a short note that it is not a licensed professional.
OpenAI repeats the point across its products. Its guide to using ChatGPT says “Legal, financial, medical, security, and other high-stakes decisions require appropriate expert review,” and the Finances page says ChatGPT “is not a replacement for professional financial advice.” Add the agent’s training against high-stakes financial actions and the directory’s ban on plugins that execute trades, and the pattern is consistent: OpenAI’s own products help with the research and leave the order to you and your broker.
The brokers that connect to ChatGPT say the same about who carries the risk. Interactive Brokers tells clients who let AI trade for them that they “retain responsibility for all trading activity.”
ChatGPT is a general assistant: superb with words and code, with the market, the account and the safety rails left for you to wire up. An AI trading agent starts from the other end, with the market data, the account and the rules already in place. Ethy is one, the AI trading agent you ask before you trade, and it covers the jobs this page found ChatGPT can’t do alone:
It handles the order too. From a chat, Ethy trades Hyperliquid perpetuals, futures with no expiry date that cover crypto as well as stocks, indices, commodities and FX listed on trade.xyz, and it runs spot swaps. Leverage on perps magnifies losses as much as gains. A setting decides whether each chat order first shows you a confirmation card. An automation trades by itself once you activate it, or runs in paper mode on live prices with no money at stake, and on perps its stop loss and take profit sit on Hyperliquid, where they trigger even when Ethy is closed. Ethy is software for research and analysis, and it doesn’t give investment advice.
ChatGPT stays the better choice for the work that isn’t a market: long documents, writing, software. For the trade itself, from the first question to the order and the stop, a tool built for the job needs far less wiring. Our Ethy vs ChatGPT comparison goes through a trading day job by job.
Yes, in two ways. Add Robinhood’s Trading MCP to ChatGPT in developer mode; it can then read your Robinhood accounts and trade in a separate agentic account you fund. Or use Robinhood Agents inside the Robinhood app and pick an OpenAI model such as GPT-6 Luna. Either way, check the trade approval setting before the first order.
Only through an account you connect. Robinhood’s Trading MCP includes crypto tools for its agentic account, and code that ChatGPT writes can call an exchange’s API. OpenAI’s directory rules bar plugins that execute crypto transfers or trades. Crypto trades around the clock, so keep a real stop order at the exchange rather than in a chat.
Not from OpenAI. Its rules bar plugins in the ChatGPT directory from executing trades, so a “ChatGPT trading bot” sold on Telegram or for binary options is someone else’s product using the name. Judge it like any third-party trading software, test it on paper first and be wary of anyone who promises profits.
Yes. Describe the rules and ChatGPT or Codex will write the code for a broker’s API, such as Alpaca’s. Your program then places the orders, not ChatGPT, so run it on a paper account first, keep the API keys out of the chat and read every line that sends an order.
Not well. Every action goes through a model turn, ChatGPT Work stops to confirm consequential actions, and a chat can’t keep a stop in place. Day trading needs orders that fire at the venue the moment a price prints. Use ChatGPT to prepare the plan and review the day, and leave execution to your broker’s own order types.
Not by itself. Without search it answers from training data that ends at a cutoff, and with search it reads whatever a page says, which may be delayed or old. Ask for the source and time of every price it gives you, and check anything you act on against your own chart or your broker.
Trading involves risk, including the loss of your money. Ethy AI is a research and trading tool and does not give investment advice or recommendations. Backtests are simulations on historical data and do not predict future results.
Deploy your AI trading agent in one click.
