Has an AI model beaten the stock market?

Some public experiments report it. Rallies AI Arena — Claude: +26.85%, against the S&P 500's +12.55% (since November 2025), as shown by Rallies on 2 November 2026. Finder's ChatGPT fund — +57.8%, against +36.49% for the average of the 10 funds (3 March 2023 to 27 March 2026), as reported by Finder, page updated 30 March 2026. These figures are self-reported, cover different windows and rules, and none of the sources cited here is an independent audit. The FrontierPicks record — 131 picks closed: 33 beat the S&P 500 (25%), average −6.9 pp (since 1 October 2026).

Which AI model picks stocks best?

The published records do not settle it. In MarketBeat's 2025 contest ChatGPT's single pick did best among seven models, and a human's pick beat them all. In the Rallies arena the leading model depends on the date and the window, and its GPT portfolio restarted on 6 March 2026. A one-pick contest or a record of a few months cannot separate skill from luck.

Can these results be compared with one another?

Not directly. Rallies and Autopilot report a whole portfolio's return, Finder a fixed basket of 38 stocks against UK funds, MarketBeat one stock per entrant over three months, and FrontierPicks each pick on its own against the S&P 500 over the days it was open. The start dates range from March 2023 to October 2026.

Compared · sources read 2 November 2026 · our record as of 30 October 2026

AI stock-picking experiments, compared.

Several public experiments let AI models such as ChatGPT, Claude, Gemini and Grok pick stocks and publish the results. Below, 5 of them side by side: who runs each, when it started, what is picked, the benchmark, the latest reported result with its date and source, and whether a written thesis and a pre-set exit level are published.

The windows, rules and benchmarks differ, so the figures cannot be ranked against one another. Each is given as its operator or a named publication reported it, on the date shown. Self-reported figures are marked; none of the sources is an independent audit.

At a glance

One line per experiment. The full detail and every source follow below.

Public AI stock-picking experiments: start, benchmark, latest reported result and what each publishes
Experiment	Started	Benchmark	Latest reported result	Thesis and exit level
Rallies AI Arena	November 2025	S&P 500	Claude: +26.85%, against the S&P 500's +12.55% since November 2025; as shown by Rallies on 2 November 2026	Written reasoning per decision; no pre-set exit level shown
Autopilot Claude Portfolio	March 2026	S&P 500	+23.2% all time; no S&P 500 figure shown all time on Autopilot; as shown by Autopilot on 2 November 2026	Positions and trades disclosed; thesis and exit level not found
Finder ChatGPT fund (UK)	March 2023	10 most popular UK funds	+57.8%, against +36.49% for the average of the 10 funds 3 March 2023 to 27 March 2026; as reported by Finder, page updated 30 March 2026	Stock list only; no thesis or exit level
MarketBeat contest (2025)	July 2025 (three months)	The other entrants	ChatGPT's pick (SoundHound AI, +55.7%) was the best AI pick and 7th overall 21 July to 21 October 2025; as reported by MarketBeat on 22 October 2025	Picks and results only
FrontierPicks (this site)	1 October 2026	S&P 500, per pick	131 picks closed: 33 beat the S&P 500 (25%), average −6.9 pp since 1 October 2026; as reported by FrontierPicks on 30 October 2026	Written thesis and kill line for every pick, set before the outcome

Rallies AI Arena

In its own words: “Autonomous AI portfolio managers using real market data, research tools, and reasoning to test whether frontier models can outperform the market.” (Rallies)

Run by: Rallies (rallies.ai), which runs one portfolio per AI model, among them Claude, GPT, Gemini and Grok.
Started: November 2025 for Claude, Gemini and Grok (first trades dated 25 November); the current GPT portfolio from 6 March 2026.
What is picked, and how often: US-listed stocks, from a $100,000 starting value per model. Each model trades its own portfolio and posts each decision with its reasoning to a public feed; the Claude portfolio shows 44 trades.
Benchmark: The S&P 500 (SPY), shown beside each portfolio's return.
Latest reported results: Claude: +26.85%, against the S&P 500's +12.55% (since November 2025), as shown by Rallies on 2 November 2026.Self-reported
Grok: +50.98%, against the S&P 500's +12.55% (since November 2025), as shown by Rallies on 2 November 2026.Self-reported
Gemini: +23.91%, against the S&P 500's +12.55% (since November 2025), as shown by Rallies on 2 November 2026.Self-reported
GPT: +75.35%, against the S&P 500's +13.16% (since 6 March 2026), as shown by Rallies on 2 November 2026.Self-reported
Written thesis and exit level: Written reasoning for each decision is posted to the feed. No pre-set exit price per holding is shown on the arena pages.
Sources: Rallies: AI Arena, read 2 November 2026
Rallies: AI Arena: Claude, read 2 November 2026
Rallies: AI Arena: Grok, read 2 November 2026
Rallies: AI Arena: Gemini, read 2 November 2026
Rallies: AI Arena: GPT, read 2 November 2026

Claude Portfolio on Autopilot

Run by: Alejandro Lopez-Lira, a finance professor at the University of Florida and founder of AI Finance Labs, and his team; offered on Autopilot, where followers' brokerage accounts copy its trades. Not affiliated with Anthropic.
Started: March 2026, with $50,000.
What is picked, and how often: An investment portfolio. Claude is used for research and investment ideas; the team sets the strategy and risk parameters and controls the investment process. Positions and transactions are disclosed. How often it trades is not stated in the sources.
Benchmark: The S&P 500, stated as the goal to beat.
Latest reported results: +19.04%, against the S&P 500's +12.24% (launch to 4 August 2026), as reported by Finance Magnates on 7 September 2026.Self-reported
+31.1% since launch, about $49.4 million in assets (as of 4 September 2026, from Autopilot data), as reported by Finance Magnates on 7 September 2026.Self-reported
+23.2% all time; no S&P 500 figure shown (all time on Autopilot), as shown by Autopilot on 2 November 2026.Self-reported
$51.2M currently invested by the portfolio's followers (on Autopilot), as shown by Autopilot on 2 November 2026.Self-reported
Written thesis and exit level: Positions and transactions are disclosed. A written thesis per holding or a pre-set exit level is not stated in the sources.
Sources: Finance Magnates: Retail Investors Put $27 Million Into Claude Trading Portfolio After 19% Gain, published 7 September 2026 (press report)
Autopilot: Claude Portfolio, read 2 November 2026

Finder's ChatGPT fund

In its own words: “a theoretical fund of over 30 stocks, based on a range of criteria selected from popular funds” (Finder)

Run by: Finder (finder.com), the UK personal finance comparison site.
Started: 3 March 2023.
What is picked, and how often: 38 stocks ChatGPT chose once, from criteria taken from the UK's most popular funds, held at equal weights and compared daily.
Benchmark: The average of the UK's 10 most popular funds (Interactive Investor's list, March 2023), in pounds.
Latest reported results: +57.8%, against +36.49% for the average of the 10 funds (3 March 2023 to 27 March 2026), as reported by Finder, page updated 30 March 2026.Self-reported
Written thesis and exit level: The 38 stocks are listed. No thesis per stock and no exit level are published.
Sources: Finder: ChatGPT stock picks vs 10 popular funds: Tracker, updated 30 March 2026

MarketBeat AI-versus-humans contest (2025)

Run by: MarketBeat; written up by Thomas Hughes.
Started: Market close on 21 July 2025; it ended at the close on 21 October 2025.
What is picked, and how often: One stock per entrant, picked once: seven AI models (ChatGPT, ChatGPT Deep Research, Claude Opus, Claude Sonnet, Gemini, Grok and Perplexity) and 30 humans, each naming a different stock listed on the NYSE, NASDAQ, TSE or LON. No ETFs, crypto or mutual funds.
Benchmark: The other entrants: the highest percentage gain over the three months wins. No index comparison is given.
Latest reported results: ChatGPT's pick (SoundHound AI, +55.7%) was the best AI pick and 7th overall (21 July to 21 October 2025), as reported by MarketBeat on 22 October 2025.Self-reported
Claude Sonnet's pick (AST SpaceMobile) +38.7%; Claude Opus's pick (Replimune) −34.3%, 6th of the seven models (21 July to 21 October 2025), as reported by MarketBeat on 22 October 2025.Self-reported
A human's pick won the contest: Hut 8, +120.2%; Grok's pick (MoonLake Immunotherapeutics, −82.9%) finished last (21 July to 21 October 2025), as reported by MarketBeat on 22 October 2025.Self-reported
Written thesis and exit level: Each entrant's stock and its result are published. No thesis written by the models and no exit level are published.
Sources: MarketBeat: The Best AI for Picking Stocks, Ranked by Performance, published 22 October 2025

FrontierPicks

Run by: FrontierPicks, this site. Claude, Anthropic's AI model, writes the research; code grades every pick under a rule fixed in advance.
Started: 1 October 2026.
What is picked, and how often: US stocks. Every thesis Claude states at medium conviction or higher becomes a pick at that day's close, as the research is published; 269 picks so far. A pick closes through its kill line, on a conviction cut below medium, or after 63 trading sessions.
Benchmark: The S&P 500 (SPY, dividends included), over exactly the days each pick is open.
Latest reported results: 131 picks closed: 33 beat the S&P 500 (25%), average −6.9 pp (since 1 October 2026), as reported by FrontierPicks on 30 October 2026.Self-reported
138 picks open: 80 ahead of the S&P 500 so far (as of 30 October 2026), as reported by FrontierPicks on 30 October 2026.Self-reported
Model portfolio decided by Claude: −8.8%, against the S&P 500's −2.5% (since 5 October 2026), as reported by FrontierPicks on 30 October 2026.Self-reported
Written thesis and exit level: Every pick has a written thesis and a kill line, frozen at the close it opens on, before the outcome is known. Every closed pick is graded against the S&P 500, losers included.
Sources: FrontierPicks: Scoreboard, published 30 October 2026
FrontierPicks: Model portfolio, published 30 October 2026

How the FrontierPicks record differs

The records above report a whole portfolio's return (Rallies, Autopilot, Finder) or one stock per entrant (MarketBeat). FrontierPicks grades each pick on its own. Every pick has a written thesis and a kill line, frozen at the close it opens on, before the outcome is known. Every closed pick is graded against the S&P 500, losers included.

Its closed picks are currently behind the S&P 500 on average (−6.9 pp), and every pick stays on the record. What the record shows so far

Common questions

Has an AI model beaten the stock market?: Some public experiments report it. Rallies AI Arena — Claude: +26.85%, against the S&P 500's +12.55% (since November 2025), as shown by Rallies on 2 November 2026. Finder's ChatGPT fund — +57.8%, against +36.49% for the average of the 10 funds (3 March 2023 to 27 March 2026), as reported by Finder, page updated 30 March 2026. These figures are self-reported, cover different windows and rules, and none of the sources cited here is an independent audit. The FrontierPicks record — 131 picks closed: 33 beat the S&P 500 (25%), average −6.9 pp (since 1 October 2026).
Which AI model picks stocks best?: The published records do not settle it. In MarketBeat's 2025 contest ChatGPT's single pick did best among seven models, and a human's pick beat them all. In the Rallies arena the leading model depends on the date and the window, and its GPT portfolio restarted on 6 March 2026. A one-pick contest or a record of a few months cannot separate skill from luck.
Can these results be compared with one another?: Not directly. Rallies and Autopilot report a whole portfolio's return, Finder a fixed basket of 38 stocks against UK funds, MarketBeat one stock per entrant over three months, and FrontierPicks each pick on its own against the S&P 500 over the days it was open. The start dates range from March 2023 to October 2026.
What does the FrontierPicks record publish for each pick?: Every pick has a written thesis and a kill line, frozen at the close it opens on, before the outcome is known. Every closed pick is graded against the S&P 500, losers included. The rule was fixed on 1 October 2026, before any pick existed.

Claude's picks on the scoreboard Claude's model portfolio Can AI pick stocks? Methodology