The only public test where language models traded real crypto is done, and the scoreboard beats any demo.
Nof1's Alpha Arena, Oct 17 to Nov 3, 2025: six models, 10,000 USD each, perps on Bitcoin, Ethereum, Solana, BNB, Dogecoin and $XRP . Four of six ended in the red, ChatGPT down 63%, win rates between 25 and 30% (Protos, Nov 4, 2025). Three ways it gets misread:
1. Taking the two winners as proof LLMs can trade, when Qwen3 Max netted 2,232 USD on one stake, one fortnight, one regime, so count it as a single sample and read the losers' numbers before the winner's.
2. Letting the model fire orders at its own pace, which is how Qwen paid 1,654 USD in fees and Nof1 wrote that P&L was dominated by trading costs, so split the roles: the model proposes, a rule you wrote decides, and nothing with an expected move under the 0.1% spot fee gets sent.
3. Trusting a paper run, when an arXiv study (Sept 2026) had the same Qwen agent go +10.1% on paper and +3.7% live, so test on a tiny live stake with a Binance API key that is read-only, withdrawals off and IP-restricted, and have Claude or ChatGPT write the trade note on $XRP while you press the button.
Two weeks of real money settled one thing: the model is the analyst, you are the trader.
#AutomateYourCrypto #XRP #AITrading #AIAgents
Nof1's Alpha Arena, Oct 17 to Nov 3, 2025: six models, 10,000 USD each, perps on Bitcoin, Ethereum, Solana, BNB, Dogecoin and $XRP . Four of six ended in the red, ChatGPT down 63%, win rates between 25 and 30% (Protos, Nov 4, 2025). Three ways it gets misread:
1. Taking the two winners as proof LLMs can trade, when Qwen3 Max netted 2,232 USD on one stake, one fortnight, one regime, so count it as a single sample and read the losers' numbers before the winner's.
2. Letting the model fire orders at its own pace, which is how Qwen paid 1,654 USD in fees and Nof1 wrote that P&L was dominated by trading costs, so split the roles: the model proposes, a rule you wrote decides, and nothing with an expected move under the 0.1% spot fee gets sent.
3. Trusting a paper run, when an arXiv study (Sept 2026) had the same Qwen agent go +10.1% on paper and +3.7% live, so test on a tiny live stake with a Binance API key that is read-only, withdrawals off and IP-restricted, and have Claude or ChatGPT write the trade note on $XRP while you press the button.
Two weeks of real money settled one thing: the model is the analyst, you are the trader.
#AutomateYourCrypto #XRP #AITrading #AIAgents