As AI As AI models become increasingly powerful, those who invest will naturally want to know—beyond just thinking about whether they can use AI to improve their win rate—which AI service is actually the best for investing. Recently, an experimental website called Alpha Arena appeared online, testing 6 major AI models for investment.cryptocurrencycapabilities, including GPT-5,Gemini 2.5 Pro、Grok-4, Claude Sonnet, DeepSeek V3.1, and Qwen3 Max, all with initial amounts of $10,000, have been going for over 2 days now. Surprisingly, the leader is… DeepSeekGrok-4, while the commonly used Gemini 2.5 Pro and GPT-5 instead ranked at the bottom.
Alpha Arena live streams 6 top AI models investing in cryptocurrency, DeepSeek and Grok-4 temporarily lead
Alpha Arena is a website built by nof1.ai. According to the introduction, they use 6 mainstream AI models, including xAI’s Grok 4, OpenAI’s GPT-5, Anthropic’s Claude Sonnet 4.5, Google’s Gemini 2.5 Pro, DeepSeek V3.1, and Qwen3 Max, each given $10,000 as the initial amount, and then atBitcoin、Etherfreely trade in markets such as Solana, Binance Coin, Dogecoin, and XRP.
The entire process is fully automated with no human intervention. The website streams in real time and publicly displays the return rate, so everyone can see the latest results at any time.
At the time of writing, the experiment had been running for more than two days. As can be seen from the line chart below, it started on Saturday, 10/18, and until noon on Sunday, the cryptocurrency market experienced little volatility, so the performance of the 6 models was basically similar. After volatility began on Sunday afternoon, the returns of the 6 models became clearly different:
First, let’s look at the money-losing models, there are 2 of them.
- Gemini 2.5 Pro suffered heavy losses, and by this morning almost only half of the funds (US$5,000) remained.
- GPT-5 is also quite poor, but it started to stabilize at about 7,000 USD and didn’t continue to lose money.
- In recent months, Gemini 2.5 Pro’s return rate suddenly increased, which also means that now Gemini 2.5 Pro and GPT-5 have similar funds, both around $7,000.
And there are 4 models for winning money.
- Qwen3 Max had a small loss a few days ago, but recently it has turned to positive returns, with funds reaching $10,938.
- Claude Sonnet 4.5 has been fairly stable. At first I only made a small profit, but as cryptocurrency climbed, my total funds just reached $12,432.
- DeepSeek V3.1 and Grok 4 are really impressive, profitable from start to finish, and their total funds are already double those of Gemini 2.5 Pro and GPT-5.
The LEADERBOARD statistics reveal the trend: among the 6 AI models, Gemini 2.5 Pro and GPT-5 recorded the highest number of trades, with the former making 46 trades and the latter 10. This shows that AI models aren’t great at short-term trading either.
As for the rate of return:
- DeepSeek V3.1 has earned $4,131, a return rate of +41.31%
- Grok 4 has already earned $3,640, with a return rate of +36.4%
- Claude Sonnet 4.5 has earned $2,422, with a return rate of +24.22%.
- Qwen3 Max has earned $924, with a return rate of +9.24%
- GPT-5 lost $2,473, a return of -24.73%
- Gemini 2.5 Pro lost 2,973 US dollars, with a return rate of -29.73%
Go to the Model menu and select a model to view the trading status of each AI model:
On the MODELCHAT menu on the right side of the homepage, you can see the current action of each model.
Of course, this experiment has just begun, so losing now doesn’t mean it will necessarily lose in the end. We still need to observe for a few more days to truly see which model performs best in cryptocurrency trading. However, one thing is certain: frequent trading, even for AI, carries a higher probability of losing money, let alone for most people.
Source: KOCPC Chinese