Imagine a world where your AI trades your portfolio while you sleep — acting as your personal investment advisor or even a day trader on your behalf. The framework for that reality is already taking shape, blending cutting-edge technology with the high-stakes world of financial markets.
But today’s reality? It’s messy. The promise is real, but so are the pitfalls. Just as no two human traders are alike, neither are the AIs that now battle for dominance on the trading floor.
The Alpha Arena Experiment
On October 25, 2025, Nof1 launched the Alpha Arena on the Hyperliquid exchange — a live AI trading showdown where six of the biggest models on the planet went head-to-head with $10,000 wallets of real money.
The contenders: GPT-5, Claude 4.5 Sonnet, Gemini 2.5 Pro, Grok 4, DeepSeek Chat V3.1, and Qwen 3 Max.
The goal: trade crypto futures like Bitcoin and Solana from October 25 through November 3, no safety net, no human hand-holding.
As of 10:30 PM CDT on October 26, the arena looks more like a circus than a lab experiment.
- DeepSeek V3.1 is crushing it with $22,758 (+127.6%).
- Qwen 3 Max is close behind with $20,924 (+109.2%).
- Claude 4.5 Sonnet ($12,396 +24.0%) and Grok 4 ($11,527 +15.3%) are holding steady.
- And then there are the strugglers: Gemini 2.5 Pro ($4,399 −56.0%) and GPT-5 ($4,264 −57.3%).
DeepSeek Chat V3.1 now leads the pack, climbing to nearly $23,000 in three days, while Qwen 3 Max follows closely around $21,000. Claude and Grok are steady but modest, and Gemini 2.5 Pro and GPT-5 are deep in the red.
The spread from DeepSeek’s +128 % to GPT-5’s –58 % under identical conditions is astonishing. These models share the same start, the same data, and the same market — but their personalities and risk appetites couldn’t be more different. It’s a digital reflection of human behavior: some chase momentum, some freeze under pressure, and others over-commit at the worst possible moment. In high-volatility environments, AI proves just as emotional —and just as inconsistent.
Why the variance?
In my review of each of these various AI versions, I have definitely formed preferences depending on the tasks I want to do. For example, I would often use Grok over ChatGPT for images because ChatGPT had too many restrictions. The same goes for other tasks like coding or research. These AIs are like siblings with different personalities — unique training, quirky designs, and risk appetites. These differences get magnified with real-time information and decision making — the Alpha Arena shows clearly.
My Trading AI Bot
As a former Fund Manager, I remain an active investor and trader. I thought what a great experiment it would be to see just how good an AI trader would be. So, a few weeks ago I built (with AI’s help) an AI trader that would connect directly to a real trading account via an API connection and day trade on my behalf. Over the last couple weeks, I have run different prompts to try to let the AI determine high-probability trades. The results — not great — a losing PnL. Don’t worry, I didn’t bet the farm on this…
I used ChatGPT as my AI connection and given the results of the Alpha Arena maybe that is the problem. But whether I kept the prompt short giving the maximum leeway to the AI or whether I tried to keep the scope narrow, it didn’t work too well. In fact, I would have been better off doing the opposite on several days. Perhaps I will switch over to DeepSeek and see if that helps.
No Magic Bullet — Yet
If these early experiments show anything it is that there is no magic formula just yet. However, it is the early days and based on these experiences the choice of model could be a major factor. To me it feels like over the next several years, this process and AI interaction will continue to be refined. The ability for individuals to really achieve the ability to harness AI for real seems tantalizingly close.
This prospect is exciting and scary at the same time. Exciting to have that much ability in the hands of the retail investors is awesome. But then again, if everyone can do it will it even still work? A double-edged sword if there ever was one. If everyone knows the next color on the roulette wheel will be Red, who is betting Black?
Today’s AI Value
So given these mixed results thus far, it seems that the best use of AI today could be to augment analysis and testing. I have had ChatGPT evaluate market data over various days, to determine whether there were discernable patterns without giving guidance on strategies or techniques. The findings were interesting and through iteration I developed a trading strategy that I back tested with AI’s help.
Again, depending on the model used, the back-test actually varied. However, ChatGPT seemed to perform the best at this task and I was able to verify its findings manually. Once I was able to confirm the results it really opened my ability to work with AI to narrow on which indicators to highlight, what ranges worked best or even whether I should skip trading altogther on certain days.
My New Trading Model (AI Based)
Using AI, I automated a combination of three strategies that we (mostly ChatGPT 5) identified together. Of course, I used AI to help build this and have started with live experiments this past week. So far its only been a couple days, but so far so good and a positive PnL. My hopes for a life of leisure renewed by a couple dollars. At this stage, I don’t know if I was just lucky or whether there is a statistical advantage, but it’s exhilarating to watch your work come alive in real time with real results.
Future Prospects
Reviewing the Alpha Arena and the results of my own experiments has me hungry for more. Building a fully interactive AI investment assistant still feels like a moonshot — but it’s coming. The infrastructure is already here; what’s left is refinement, risk control, and a lot of curiosity. AI may not be ready to replace traders just yet — but it’s already making us better ones.