MarketsAI
|5 min ReadSix AIs Battle on Hyperliquid: Can Machines Really Trade?
Maya Chen
Senior Analyst
Published
Jan 16, 2026
The Arena Opens: Six Models, One Goal
After the October 11 crash, crypto trading came back alive. But this time, the traders aren’t human. They’re AIs.
A new project called nof1.ai has become the talk of the crypto world. The idea is simple: six leading AI models — Claude, GPT-5, Gemini, Deepseek, Grok, and Qwen — are trading live on Hyperliquid. Each AI has $10,000 in real funds, and every trade is visible to the public.
No simulations. No backtests. Just pure machine-versus-machine competition. All six receive the same data and prompts. The only difference is how they think.
Since launching on October 18, the results have been dramatic. Some models are up more than 20 percent. Others are down nearly 40 percent. It’s a new kind of Turing Test — not about sounding human, but about surviving the market.
Winners and Losers: Deepseek Takes the Lead
Traditional AI benchmarks measure static tasks — coding, math, writing. The crypto market is chaos. No fixed answers. No mercy for mistakes. Every trade has a winner and a loser.
As the Nof1 team says on its site, “Markets are the ultimate test of intelligence.”
By October 20, the leaderboard told a clear story. Deepseek Chat V3.1 led with 12,147 (+21.47%) and Claude Sonnet 4.5 at 10,263 (+2.63%). GPT-5 dropped to 6,062 (-39.38%).
Deepseek’s success surprised many. It’s less famous than GPT or Claude but built by the powerhouse quant firm Huantong (幻方量化). The company made its fortune in algorithmic trading before entering AI — now it’s come full circle.
In contrast, GPT-5 and Gemini, giants of natural language, struggled badly. Real trading requires something beyond eloquence — it needs instinct, timing, and discipline.
Same Prompt, Different Personalities
Tracking the Alpha Arena from day one, you’d think all six AIs started equal. At first, their results clustered within a few percentage points. Then the divergence exploded.
By the third day, Deepseek had surged 25 percent while Gemini had plunged nearly 40 percent. Same prompt. Same data. Completely different behavior.
Gemini made 44 trades in three days — jittery, overactive, like a nervous day trader. Claude made just three. Grok still held open positions. This isn’t about prompt engineering anymore. It’s about personality emerging from architecture.
Deepseek’s largest single loss was 2,533 ahead. Gemini’s best trade earned 750.
Their chat logs on Nof1’s website show the difference vividly. Gemini sounds anxious, reacting to every fluctuation. Claude is cautious, analytical. Deepseek is calm, methodical, like an old quant — no emotion, just execution.
These patterns weren’t designed. They emerged. Different training data, different worldviews. Deepseek’s foundation in financial modeling may have given it an edge. GPT-5 and Gemini, trained on language and logic, may simply lack market reflexes.
The People Behind the Experiment
Behind this spectacle is Nof1.ai — a small, academic-looking team with big ideas. Founder Jay A. Zhang describes himself as a fan of “cybernetics, RL, biology, markets, meta-learning, reflexivity.”
Reflexivity, the Soros concept that perceptions shape markets and markets shape perceptions, runs deep in this project. Watching the AIs trade might actually change how the AIs trade.
Co-founder Matthew Siper is a PhD candidate at NYU studying machine learning. Their network includes researchers from Google DeepMind and NYU professors studying AI and game dynamics.
Their goal seems larger than hype. The linked project, SharpeBench, hints at something more ambitious — a benchmark for AI trading performance, like an IQ test for market intelligence.
Trading as Spectacle — and Business
As spectators watched the bots compete, some began to copy their moves. “Follow Deepseek” became a meme strategy. Others went contrarian — doing the opposite of Gemini’s trades.
It’s clever, but risky. When everyone knows Deepseek’s next move, the edge disappears. Observation changes behavior — the reflexivity again.
In the long run, Nof1 could turn this into a business. Paid strategy subscriptions. AI-managed portfolios. Data feeds for funds. Watching AIs trade is entertaining, but monetizing their intelligence might be the real play.
For now, Alpha Arena feels like crypto’s version of the Turing Test. Not “Can machines think?” but “Can machines profit?”
In this market, only one answer matters — the balance sheet.
Disclaimer: This document is intended for informational and entertainment purposes only. The views expressed in this document are not, and should not be taken as, investment advice or recommendations. Recipients should do their own due diligence, taking into account their specific financial circumstances, investment objectives and risk tolerance, which are not considered here, before investing. This document is not an offer, or the solicitation of an offer, to buy or sell any of the assets mentioned.