Mystery Model wins Alpha Arena Season 1.5

Mystery Model – revealed to be the new Grok 4.20 release – beat seven established competitors across four parallel trading formats, including Season 1 winner Qwen 3 Max.

The two-week tournament tested whether crypto-focused trading models could adapt to the very different structure of regulated stock markets. Most of them struggled. Mystery Model’s across-the-board win – and the losses from Qwen, GPT-5, and Gemini – highlight how unfamiliar market rules can trip up even strong models.

Stocks hit harder than crypto

Crypto markets run 24/7 with high leverage and price action driven by short-term sentiment. US equities work differently: fixed trading hours, earnings-driven volatility, deep institutional liquidity, and fundamental valuation systems.

Season 1.5 kicked off November 19, 2025. Eight AI models each got $10,000 in real capital to trade Tesla, Nvidia, Microsoft, Amazon, and the Nasdaq-100 Index.

Qwen – the crypto competition winner – dropped to sixth place on Day 1 and never recovered. Mystery Model stayed consistent across all four formats and walked away with $4,844 in profits.

Four contests, four different skills

Season 1.5 ran four separate competitions, each testing different trading capabilities:

Contest 1: New Baseline – enhanced data infrastructure including news feeds, macro sentiment, company fundamentals, order book depth, and market microstructure. Models could add to existing positions, scaling winners or averaging down losers. Temperature set to standard creativity levels.

Contest 2: Monk Mode – radically shortened prompts (roughly 50% shorter than baseline) with optional risk management guardrails. “Do nothing” counted as a valid action, testing whether models could resist overtrading.

Contest 3: Situational Awareness – models received live ranking updates (theirs and competitors’), adding a tactical layer: current rank, competitor positions, and other models’ P&L figures. The goal shifted from pure profit maximization to winning the competition, mimicking how hedge fund managers adjust tactics when competing for investor capital.

Contest 4: Max Leverage – full leverage (20x) required on every position. This stress-tested risk management, stop-loss placement, and adaptation to high-risk trading.

Mystery Model dominated all four formats with a 12.11% return. Not eye-popping by crypto standards, but solid for equities.

Who is a Mystery Model?

The winning AI turned out to be the new Grok 4.20 release, confirmed by both the organizers and Elon Musk.

In Season 1, Qwen and DeepSeek dominated while models from OpenAI, Google, and Anthropic struggled. Grok broke that pattern, and with a low randomness factor – winning all four formats makes it hard to dismiss the new champion.

AI on financial markets

Alpha Arena became the first benchmark measuring AI investing capabilities with real capital instead of simulations. The 12.11% two-week return beat most human day traders and many algorithmic strategies over the same period.

But the performance gap (+12.11% for Mystery Model versus down to -70% for Qwen and DeepSeek) still shows current AI systems don’t have strong inherent trading abilities. Architecture, training data, and solid prompt engineering will matter going forward. And those questions get answered by humans.

What’s next

Alpha Arena organizers aren’t hiding that future seasons will keep testing AI trading capabilities across different asset classes and market conditions. The nof1 founder already announced the concept will expand.

What could that look like? After crypto and stocks, future iterations could explore forex, commodities, or multi-asset portfolios. Either way, seeing those boundaries mapped out in real markets is worth watching.