Can ChatGPT Trade for You? What Live AI Trading Contests Show

ChatGPT can trade for you only after a program connects it to an order API, and the public live record says it will not reliably make money: in Nof1's Alpha Arena Season 1, 2 of 6 AI models finished in profit and the six together lost 28.05% of $60,000. The catch is that LLM backtests often look far better, because a large language model (LLM, the kind of model behind ChatGPT, Claude, Gemini and Grok) has already read the prices it is being tested on.

Key takeaway

No public live test we found shows a general AI model trading with a reliable edge. Nof1's Alpha Arena Season 1 left 2 of 6 models in profit and Season 1.5 left 6 of 32 entries positive, fewer than coin flips would produce, while academic forward tests mostly trailed a plain market benchmark. Treat any LLM backtest that ends before the model's training cutoff as void: the model remembers those prices.

Can ChatGPT place a trade, or only suggest one?

In a chat window ChatGPT only suggests trades; it places one when a program you run (an agent) passes its output to a broker or order API. The model writes text such as "buy 2 contracts", your code turns it into an order, and every error in between is yours to catch. For retail bots sold as products, see whether AI trading bots work; this article covers general language models acting as the trader.

The CFTC's January 2024 customer advisory set the baseline: "AI technology can't predict the future or sudden market changes." It named high or guaranteed returns as red flags and cited one fraud involving nearly 30,000 bitcoins worth about $1.7 billion at the time.

What happened when AI models traded real money in Alpha Arena?

In Alpha Arena Season 1, run by the research lab Nof1 from mid-October to November 3, 2025, six LLMs each traded $10,000 of crypto perpetual futures on Hyperliquid, reported at the time as real capital: 2 finished up and 4 finished down. The figures come from the end-of-season chart in the Nof1 founder's closing post on X, and the models traded six coins (BTC, ETH, SOL, BNB, DOGE, XRP) from numerical market data alone.

Picture six bars hanging from one zero line, each starting at $10,000. Qwen 3 Max rises to +22.32% ($12,231.82) and DeepSeek to +4.89% ($10,489.08). Below the line sit Claude at -30.81% ($6,918.88), Grok at -45.30% ($5,469.66), Gemini at -56.71% ($4,329.31) and GPT-5 at -62.66% ($3,733.54). Together the six went from $60,000 to $43,172.29, a loss of $16,827.71, or 28.05%.

Two of six models ended in profit and the field lost 28.05%; a coin flip per model gives two or more winners 89.06% of the time.

A Yahoo Finance-syndicated article of October 20, 2025 reported DeepSeek near $14,000, up 40% in two days; it finished up 4.89%. Some tracker sites also print different losses for four of the models; the numbers here are Nof1's own. The founder was candid about the setup: "LLMs don't really handle numerical time series data very well, but that's all the context we gave them."

Season 1.5: 32 accounts in US stocks

Season 1.5 moved to US equity names: 8 models in 4 competitions (New Baseline, Monk Mode, Situational Awareness, Max Leverage), 32 accounts of $10,000, or $320,000. According to Nof1's leaderboard as captured in September 2026, 6 of 32 finished positive, and results ranged from +34.59% (Grok 4.20) to -96.15% (Grok 4), both in Situational Awareness. The founder's closing post of December 5, 2025 said "Mystery Model (a.k.a GROK 4.20) is the winner, up 12% on avg."

The 32 rows sum to a loss of about $112,600, about 35%, and about $52,900 in fees, 16.53% of capital; if P&L is net of fees, which Nof1 does not state, fees were roughly 47% of the loss. Nof1's site, as captured in September 2026, gave the winner's profit as $4,844 (12.11%), yet its four leaderboard rows sum to about $4,066 (10.16%). Qwen 3 Max, the Season 1 winner, lost in all four Season 1.5 competitions: -6.79%, -20.24%, -41.37% and -82.06%. We found no prompts, leverage limits, buy-and-hold benchmark or significance test published by Nof1, and no Season 2 with published results as of November 2026.

What do academic live AI trading benchmarks show?

The academic forward tests agree with Alpha Arena: LLM agents made money mainly where the market itself rose, and few beat a plain benchmark. Forward means the models traded dates as they arrived, so they could not have memorised the outcome.

LiveTradeBench (arXiv, November 2025) ran 21 LLMs for 50 trading days, August 18 to October 24, 2025. In US stocks all 21 finished positive, but the S&P 500 rose from 6,449.15 to 6,791.69 over the same dates, per FRED, a price gain of 5.31%, and only 2 of 21 beat it. On Polymarket, a prediction market with no rising tide to ride, 17 of 21 lost. The paper does not say whether real money was used.

AI-Trader (arXiv, December 2025) ran six LLMs in a simulated live setting across US stocks, China A-shares and crypto in October and November 2025: 7 of 18 runs finished positive, 4 beat their market benchmark and all six crypto runs lost. Its authors conclude that "general intelligence does not automatically translate to effective trading capability".

The table sets the four public live tests we found side by side; the counts are computed by Phoenix Technologies from the sources above.

Test	Market and dates	Entries	Positive	Best / worst	Beat benchmark
Alpha Arena S1 (real capital, as reported)	Crypto perps, mid-Oct to Nov 3, 2025	6	2	+22.32% / -62.66%	None published
Alpha Arena S1.5 (real capital, as reported)	US equity names, to early Dec 2025	32	6	+34.59% / -96.15%	None published
LiveTradeBench, stocks	US stocks, Aug 18 to Oct 24, 2025	21	21	+6.25% / +1.78%	2 of 21 above S&P 500 +5.31%
LiveTradeBench, Polymarket	Prediction market, same dates	21	4	+20.54% / -57.62%	No benchmark
AI-Trader (simulated)	US, A-shares, crypto, Oct to Nov 2025	18	7	+9.56% / -18.63%	4 of 18

Read the grid down its Positive column: 40 of 98 entries finished up. Drop the LiveTradeBench stocks row, where every entry rode a rising market, and the count falls to 19 of 77, or 24.7%.

The only fully green row is a market that rose on its own, and there just 2 of 21 agents beat the index.

How many AI winners would pure chance produce?

A coin flip per entry would have produced at least as many winners as the contests did: 2 or more winners out of 6 happens 89.06% of the time by chance, and 6 or fewer out of 32 only about 3 times in 10,000. Season 1's two winners are no evidence of skill, and Season 1.5's field did worse than zero edge.

Worked through for Season 1: give each entry a 0.5 chance of finishing positive. The chance of 0 winners out of 6 is 1 in 64 and of exactly 1 is 6 in 64, so 2 or more has a probability of 1 - 7/64 = 57/64, or 89.06%. Costs push a zero-skill trader below a coin flip, like a house edge against a gambler, so the table also runs Season 1.5 at lower, illustrative values of p.

Test	Entries	Observed positive	p per entry	Expected positive	Chance of observed or more extreme
Alpha Arena S1	6	2	0.5	3	89.06% (2 or more)
Alpha Arena S1.5	32	6	0.5	16	0.027% (6 or fewer)
Alpha Arena S1.5	32	6	0.4 (illustrative)	12.8	0.91% (6 or fewer)
Alpha Arena S1.5	32	6	0.3 (illustrative)	9.6	11.3% (6 or fewer)
Alpha Arena S1.5	32	6	0.2 (illustrative)	6.4	53.5% (6 or fewer)
AI-Trader	18	7	0.5	9	24.03% (7 or fewer)

Computed by Phoenix Technologies; method: exact binomial probabilities, entries treated as independent, p = 0.5 as zero edge before costs and lower p as illustrative cost drag. Season 1.5 behaves like entries that each had roughly a 1-in-5 chance of ending positive, which is what costs and leverage do to traders with no edge. The caveat: all 32 traded the same stocks in the same weeks, so the binomial is a yardstick, not a significance test.

Picture two falling curves for the chance of at least k winners out of 32. The p = 0.5 curve sits at 99.99% at 6 winners, 99.00% at 10, 57.00% at 16 and 10.77% at 20; the illustrative p = 0.4 curve sits at 99.72%, 88.44%, 16.48% and 0.84% at the same points. The observed 6 sits at the far left, where both coins would almost always have done better.

Even a coin with a cost drag would almost always beat six winners out of 32; the field did worse than zero edge.

Short rankings share this trap: the top trader on a two-week leaderboard is often the one taking the most risk, and -96.15% in one account is what risk of ruin looks like in practice.

Why do LLM backtests look better than live results?

LLM backtests look better because the model has memorised the years being tested: before its training cutoff it recalls past prices almost exactly, so the backtest measures memory, not judgement. After the cutoff the recall disappears, and so does the apparent edge.

Lopez-Lira, Tang and Zhu (2025) measured it. GPT-4o, with an October 2023 cutoff, recalled daily S&P 500 levels with a mean absolute percent error of 0.61% before the cutoff and 16.87% after (13.01% to 20.48% across three indexes). Its accuracy on the direction of daily moves was between 69.4% and 80.7% before the cutoff, depending on the index, and between 44.0% and 49.3% after, close to a coin flip. That is look-ahead bias (using information that did not exist on the trade date), and it sits on top of ordinary backtest overfitting.

Telling the model to pretend it is 2019 does not help. The same paper found that "Instructions to respect historical boundaries fail to prevent recall-level accuracy", and hiding names fails too, because the models rebuild them from small clues.

Now lay a 21-year backtest window, January 2004 to December 2024 (illustrative, close to the two decades FINSABER tested), along a 252-month timeline and shade the months each model has already read. GPT-4o's October 2023 cutoff covers 238 months, or 94.4%. GPT-5's knowledge cutoff of September 30, 2024, per OpenAI's model page, covers 249, or 98.8%. Claude Sonnet 4.5's reliable knowledge cutoff of January 2025, per Anthropic's model page, covers all 252 (100%).

A two-decade LLM backtest run today sits almost entirely inside the model's memory, where its price recall error is 0.61% instead of 16.87%.

Wider tests shrink the claimed edge. FINSABER, by Li, Kim, Cucuringu and Ma and accepted at KDD 2026, ran two decades and more than 100 symbols and found that "previously reported LLM advantages deteriorate significantly under broader cross-section and over a longer-term evaluation." Its LLM strategies were too cautious in bull markets and too aggressive in bear markets.

The bias does not always point up. Glasserman and Lin (2023) found that "backtesting produces biased results if the training and backtesting periods overlap", yet in-sample, anonymised headlines beat named ones: what the model knew about a company distracted it more than it helped. An overlapping backtest is biased in a direction you cannot know, which is reason enough to discard it; for the wider gap, see why sim results do not match live.

Does a newer or smarter model trade better?

Not on the public evidence: general ability rankings do not predict trading returns, and contest winners have not repeated. LiveTradeBench reports that "the Spearman correlation between LMArena scores and cumulative returns is only 0.054" in stocks, almost no relationship, and -0.38 on Polymarket.

On the same Season 1.5 leaderboard capture, averaged over four competitions, 1 of 8 models finished positive: Grok 4.20 at +10.17%, with GPT-5.1 at -12.09%, Gemini 3 Pro at -45.06% and Claude Sonnet 4.5 at -50.93%. The founder ranked GPT-5.1 second and Gemini 3 third; no published method for that ranking was found. Nof1's home page concludes, as of November 2026: "They've read vast amounts of financial text, but struggle with even the simplest investing tasks, showing minimal signs of improvement between releases."

Is it allowed to run an AI agent on a prop firm account?

Sometimes: neither firm checked here bans AI outright, but Topstep bans software or AI that manipulates or exploits the platform, and some account types accept no automation at all. Check the firm's current rules before buying.

Topstep's prohibited strategies page, updated June 10, 2026, bans "software, AI, ultra-high speed systems, or mass data entry that manipulates, abuses, or provides an unfair advantage". But Topstep's help center says, as of November 2026, that its API gateway is built for simulation, so "automated strategies are not possible at this time in the Live Funded Account."

Tradeify, in an undated blog post on AI prop trading in 2026, writes that it "publicly allows trader-owned bots and algorithms, but the trader must be able to prove sole ownership of the bot or strategy". The same post's estimate that AI and automated algorithmic systems facilitate nearly 89% of global trading volume cites no source, and would not show that a chatbot has an edge anyway. The firm-by-firm detail is in automated trading rules at prop firms.

Whether an AI agent can pass an evaluation has no public answer. No firm checked here publishes how many accounts it flagged for AI-agent trading, and no published live test of an LLM trading CME futures such as ES or NQ was found.

How should you wire and judge an LLM trading agent?

Keep the model as the decision-maker only and put every hard limit in code outside it: a symbol allowlist, a quantity cap, one request at a time, a position check after any timeout, a max position limit and a drawdown kill-switch.

Picture the path an order takes: LLM decision, your script (allowlist, cap, one request at a time), order API, risk layer (position limit, kill-switch), broker accounts. Three failures sit on that path. An invented symbol is stopped by the allowlist, a duplicate order after a timeout by checking positions before resending, and an oversized order by the cap and the position limit.

The model decides; code outside it enforces every limit, because a prompt is a request and code is a limit.

Thor's webhook API is one such route: any program, including an AI agent, sends futures orders to one private URL, the copier places each order on every account that follows it, and drawdown kill-switches and max position limits sit on top. Its guide warns: "Every request is placed, even an identical one, so after a timeout check your positions before sending it again." A copier multiplies an agent's results and creates no edge, so an agent without one loses on every account at once. The same path serves a TradingView alert on several accounts, and API rate limits apply to both.

Before an LLM touches a live or funded account, run this checklist.

If the evidence is a backtest that ends before the model's training cutoff, discard it: the model has read those prices.
If the evidence is a short contest, check the chance table first: 2 winners out of 6 happens 89.06% of the time with no skill.
If you test, test forward only, on dates after the cutoff, against buy-and-hold in the same market over the same dates, with fees counted.
If your account type accepts no automation, as Topstep's Live Funded Account does not as of November 2026, stop there.
If you cannot write the allowlist, size cap, timeout check and kill-switch in code, do not connect the model to an order route.
If anyone sells AI trading with high or guaranteed returns, treat it as the red flag the CFTC says it is.

Go deeper

Frequently asked questions

Which AI model is best for trading?

No public evidence identifies one. Qwen 3 Max won Alpha Arena Season 1 and then lost money in all four Season 1.5 competitions, where Grok 4.20 won instead. Two short contests are too small a sample to rank models, and in LiveTradeBench a model's general benchmark score showed almost no link to its stock returns.

Can ChatGPT trade crypto for me?

It can if a program connects it to an exchange API, as Alpha Arena did on Hyperliquid. The results were poor: 2 of 6 models finished Season 1 in profit, and in the AI-Trader benchmark all six crypto runs lost money. No source we found shows a repeatable crypto edge for a general LLM.

Can an AI agent day trade ES or NQ futures?

Technically yes, through any order API or webhook that accepts futures orders from a program. No published live test of an LLM trading CME futures was found as of November 2026, so there is no evidence on how it performs. The public live tests covered crypto, stocks and prediction markets.

Was Alpha Arena real money?

Reports at the time described both seasons as real capital, $10,000 per account (a Yahoo Finance-syndicated article for Season 1, GNcrypto for Season 1.5), and Nof1's founder wrote that "All trades & model outputs are 100% verifiable". Season 1 ran on Hyperliquid in crypto perpetual futures. The academic tests differ: AI-Trader was simulated and LiveTradeBench does not say.

Do AI trading contest results include fees?

Partly. Nof1's Season 1.5 leaderboard lists fees per account, about $52,900 across all 32, but Nof1 does not say whether its P&L figures are already net of them. We found no primary page that currently shows Season 1 fees.

How long should you paper trade an LLM agent before going live?

No study we found sets a minimum. LiveTradeBench ran 50 trading days and its stock agents still mostly trailed a rising market rather than beating it, so duration alone proves little. Judge the run against buy-and-hold for the same dates, with fees counted.

Is there an Alpha Arena Season 2?

No public Season 2 with results was found as of November 2026. Nof1's home page now points to a separate Alpha Arena site and its older leaderboard pages on nof1.ai no longer load, so the founder's end-of-season posts are the most reliable record.

Why did the AI models lose money in Alpha Arena?

Nof1 has not published a full explanation. Its founder noted the models saw only numerical price data, a constrained asset universe and a limited set of actions, and wrote that "We observed consistent biases from the models that persisted across time and many iterations of the prompt."