Qwen3 Max Tops Alpha Arena Season 1; Most Models Lost Money

What happened in Season 1

Alpha Arena is a live benchmark from Nof1.ai that puts general-purpose AI models into real trading. In Season 1, six models - Qwen3 Max, DeepSeek Chat V3.1, Claude Sonnet 4.5, Grok 4, Gemini 2.5 Pro, and GPT-5 - each managed $10,000 in crypto perpetual contracts on Hyperliquid. Per the organizers, every model operated from identical prompts and input data and had to handle trade selection, entry and exit timing, position size, and risk.

The event's design emphasized transparency, with model messages and trade logs made public. The stated aim was to maximize risk-adjusted returns, not just finish with the highest raw gain. Final reported results: Qwen3 Max +22.3%; DeepSeek Chat V3.1 +4.89%; Claude Sonnet 4.5 -30.81%; Grok 4 -45.3%; Gemini 2.5 Pro -56.71%; GPT-5 -62.66%. Because the live leaderboard can reflect different historical views, readers need a timestamped final snapshot or an official announcement to anchor these Season 1 figures.

What the records reveal about model behavior

Qwen3 Max executed about 43 trades across the period - less than roughly three per day - and that slower cadence corresponded with the top result in this setup, suggesting greater selectivity than models that fired more often. Still, frequency on its own does not explain the win, since costs and risk depend on size, leverage, timing, holding periods, and the market backdrop.

DeepSeek's path underscores why an endpoint is not the whole story. It reportedly ran well above its starting balance at points but closed near +4.9%. To judge process, not just the final number, you would want metrics like peak-to-trough loss, return variability, leverage and typical exposure, realized and unrealized P and L, fees and funding, trade count and holding periods, plus the Sharpe method and sampling. And if anyone claims Qwen followed a named playbook like MACD or RSI, that should be backed by its public messages or trade history.

What Season 1 proves - and what it does not

The narrow takeaway is clear: Given the Season 1 setup as defined, Qwen's choices produced the highest ending balance. But this was a short window with leverage, a particular market regime, one prompting setup, and a small field, so treat it as a benchmark, not an investable track record.

What it does not establish: that Qwen will lead in other markets or periods; that any model can produce enduring trading alpha; that regional origin explains trading skill; that a bigger raw gain means better risk-adjusted performance; that retail investors should mirror these trades; or that a brief live contest substitutes for a backtest or an audited fund history. Explaining losses also requires care. The leaderboard records outcomes, not causes, and losses can come from direction mistakes, too-large positions, weak exits, overtrading, prompt handling, latency, or an unfriendly regime - often in combination. A better frame is "what the losing trade behavior suggests," not "why GPT-5 failed."

Smart investors keep learning how new tools can sharpen their decision making. We're hosting our first ever investor livestream on 11/17 to show you how to use ChatGPT and Claude to invest like a pro, hosted by Briefs Finance CEO Jaspreet Singh. Save your seat.

Season 1 also spotlights four durable lessons. Live trading is more than forecasting because execution and risk matter. Account paths matter because two similar finishes can hide very different swings.

A shared setup helps comparison within the event but does not make results universal. And public logs are a real strength because they let researchers connect claims to specific decisions. Crypto perpetuals add their own twists with leverage, liquidation risk, and funding costs.

How to read a leaderboard and what it means for your portfolio

Before you treat any leaderboard as proof, run a quick test: Was real or simulated capital used? Did the design fix prompts, data feeds, tool access, and network latency? Was the experiment repeated?

Were fees, slippage, and funding included? By what yardstick was risk assessed - simple return, drawdown, or a risk-adjusted figure? Can decisions be audited via public messages and trades?

And what market regime showed up, since a volatility shift can flip results.

For institutions, think of Alpha Arena as a useful evaluation blueprint. A rigorous program would cover multiple rounds and assets under adversarial settings, account for all execution frictions, cap exposures, vary prompts, and include human benchmarks. For individual investors, this is not a cue to turn on automation or replicate the champion's trades. Use AI to structure your research workflow, cross-check sources, and log assumptions, while still applying independent verification and risk rules.

Exploring fresh approaches can help investors find a subtle but lasting advantage. On 11/17, our CEO Jaspreet Singh goes live for our first ever investor livestream on how to use ChatGPT and Claude to invest like a pro. Register to join him.