An AI investing experiment tells you how specific models behaved under one set of rules, over one short window. It does not tell you which AI will reliably grow your money. The most cited example, Nof1's Alpha Arena Season 1, gave Claude, GPT-5, Gemini and three other models $10,000 each to trade crypto futures, and according to Protos four of the six finished in the red. This guide, checked against public sources as of October 2026, explains how that contest worked, what the results actually show, and a checklist for judging the next viral "I gave AI my money" video.

How an AI investing experiment works: the Alpha Arena setup

Nof1, a US startup, framed Season 1 around one question: can a large language model, with minimal guidance, act as a zero-shot systematic trader? According to its write-up, the design was:

• Models: GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, Grok 4, DeepSeek V3.1 and Qwen3-Max
• Capital: $10,000 of real money per model, with zero human intervention, on the decentralized exchange Hyperliquid
• Assets: perpetual futures on six coins (BTC, ETH, SOL, BNB, DOGE, XRP), long or short, with leverage allowed
• Inputs: numerical market data only (prices, volume, technical indicators); no news feed
• Cadence: each model was called roughly every two to three minutes to enter, hold or close positions
• Fairness: identical system prompt, data and output format for every model

Each order had to include a justification, a confidence score between 0 and 1, and an exit plan with a profit target, stop loss and invalidation condition. In one trade Nof1 published, Claude opened a Bitcoin long at 20x leverage and stuck to its plan across 443 consecutive check-ins until the target hit. In other words, this was less a test of picking good assets and more a test of following rules and managing risk in fast, leveraged trading.

Alpha Arena Season 1 results: Claude vs ChatGPT vs Gemini

Season 1 ran from mid-October 2025 until November 3, 2025 at 5 p.m. Eastern Time, a little over two weeks. The table converts the final profit and loss reported by Protos into a return on the $10,000 stake (P&L divided by $10,000). Different trackers and snapshot times can shift these figures slightly.

The final ranking hides how unstable it was. Protos reported that around October 27, Qwen3-Max and DeepSeek had roughly doubled their money and Claude and Grok were briefly profitable. Stop the clock on a different day and you get a different leaderboard. Win rates for all six sat between 25% and 30%, while activity varied widely: Gemini made 238 trades, Claude just 38.

Did each model show a consistent trading personality?

Nof1 saw fairly stable behavioral patterns across trial runs. Grok, GPT-5 and Gemini shorted far more often than the others; Claude rarely shorted at all. Gemini traded most frequently, and Qwen took the largest positions and reported the highest confidence. Notably, Nof1 found that self-reported confidence appeared decoupled from actual performance, so an AI saying it is confident is not a useful signal on its own.

Five checks before you trust any AI trading result

Whether it is Alpha Arena, a research paper or a YouTube challenge, these questions filter out most over-reading.

1. 1

   How long, and how many runs?

   A two-week or one-month single run is dominated by luck. Nof1 reported run-to-run changes in both rankings and correlations between models. Look for repeated runs across different market conditions.

2. 2

   What assets, and how much leverage?

   Leveraged crypto perpetuals magnify small price moves into large gains or losses. Results from that setting say little about buying and holding stocks, and vice versa.

3. 3

   Are fees and costs included?

   Nof1 said early test runs were dominated by trading costs because agents over-traded and fees erased small gains. In Season 1, Protos reported Qwen3-Max paid $1,654 in fees and Gemini $1,331. Make sure returns are net of fees and slippage.

4. 4

   What exactly was the prompt and data?

   Nof1 found the models highly sensitive to trivial prompt changes. When data was listed newest to oldest, several models still read it as oldest to newest. Change the prompt and you effectively change the trader.

5. 5

   Compared with what?

   A ranking between models does not show whether any of them beat simply holding Bitcoin or an index fund over the same period. Ask for a benchmark.

It also helps to know whether big scheduled events landed inside the test window. The economic calendar lists dates such as rate decisions and jobs reports so you can see what happened during an experiment; it shows schedules, not predictions.

页面引用

股票theddari.com›股票经济数据日历韩国及全球主要经济数据发布日程

The real risks of letting AI trade your money

Viral experiments feed a wave of products that promise AI will make money for you. Regulators keep flagging the same problems.

Relying only on chatbot answers

A January 2024 joint investor alert from the SEC, NASAA and FINRA warns that AI-generated information can rest on inaccurate, outdated or manipulated data, and can be made up even from accurate inputs. It adds that chatbot conversations can push people toward emotional, impulsive trades.

AI trading bot scams and AI washing

The CFTC states that AI cannot predict the future or sudden market changes. It describes a bot program that promised at least 10% a month and took more than $1.7 billion in bitcoin from at least 23,000 people. In July 2025, FINRA warned of a rise in unregistered auto-trading apps that overstate their AI (AI washing), cite unverifiable returns and ask for brokerage usernames and passwords.

Deepfake experts outside the US

The pattern is global. South Korea's Financial Supervisory Service issued a consumer alert in January 2026 about illegal stock-tip chat rooms that use AI deepfakes to impersonate real brokerage experts. If you follow AI investing chatter on Telegram, the AI Telegram trends page shows what topics are circulating, but buzz is not verification.

页面引用

趋势theddari.com›趋势Telegram · AIAI 频道实时帖子

Using AI for investing: what to delegate and what to keep

Taken together, the experiments and regulator warnings point to a practical split: general-purpose AI is lower risk as a research and checking assistant than as an autonomous trader.

Bottom line: AI trading contests are useful for revealing how models differ in risk appetite and habits. A two-week single run, leveraged futures and prompt-sensitive behavior mean they cannot tell you which chatbot to hand your savings to. Check the time frame, assets, costs, prompts and benchmark of any experiment, and treat any AI product that promises steady, outsized returns as a reason to verify its registration first.