Before you hand a language model an API key, make it trade on paper for a month and count the trades, because the public record says that count is where it dies.
Nof1's Alpha Arena, Oct 17 to Nov 3, 2025, gave six models 10,000 USD each on live perps. Four lost, ChatGPT 63% of its stake, and Nof1 said fees erased most gains. Season two, reported May 2026: eight models, the pool lost about a third, Qwen alone fired 1,418 trades. That proves a bare LLM on raw prices overtrades, not that AI is useless. Same test at home,
$TAO at about 254 USDT as the example:
1. Make a key on testnet.binance.vision (fake balance, real order book) and a read-only key on your main account for data, withdrawals off, IP restricted.
2. A Python loop pulls TAOUSDT 4h klines from /api/v3/klines into Claude or ChatGPT with one fixed prompt: JSON only, buy, sell or hold, size, one line of reasoning.
3. Act only on candle close, six decisions a day at most, and charge 0.1% per fill in your log even though testnet doesn't.
4. After 30 days, compare against holding TAO and count the trades.
Cost: nothing on Binance, cents per model call. It goes wrong when you skip the fee line and a positive curve is just the exchange getting paid. I'm watching whether
$TAO holds 250.
#AutomateYourCrypto #TAO #AITrading