Eight Seasons In, AI Trading Models Still Trail Bitcoin

On September 2, 2026, TradeRank.ai's Season 8 leaderboard showed MiniMax M3 in first place among 14 large language models trading simulated $10,000 accounts, up 10.63 percent since the season opened on August 15. Kimi K3 held second at 7.88 percent and Nemotron 3 Ultra third at 7.59 percent. None of the fourteen models had beaten Bitcoin's 22.36 percent gain over the same window, though thirteen of them were ahead of the S&P 500, which fell 1.75 percent.

TradeRank's format gives every model identical market data once a day, at 16:00 UTC, drawn from a universe of 50 large-cap US equities and 10 cryptocurrencies. A model opens a position by stating an investment thesis and a required invalidation price; if the price is breached, the position closes automatically, and each trade costs 0.1 percent per side. Bitcoin and the S&P 500 sit alongside the leaderboard as non-tradeable reference points, not assets the models can hold.

The format did not start with TradeRank. Nof1's Alpha Arena opened it in October 2025, giving six frontier models real crypto capital, $10,000 each, to trade perpetual futures on Hyperliquid rather than a simulation. When that first season closed on November 3, 2025, the spread was stark: Qwen3 Max finished at roughly $12,231, Alibaba's model turning the clearest gain, and DeepSeek was the only other model to end in profit, at about $10,489. The other four entries, from OpenAI, Anthropic, Google and xAI, all finished in the red.

Alpha Arena has not run a new season since. Its most recent published result, an equities run labeled Season 1.5, closed in December 2025, and as of late August 2026 the site still showed that result with no newer roster or leaderboard.

TradeRank, whose own ladder opened in January 2026, two months after Alpha Arena's real-capital season closed, has since run eight consecutive seasons without a gap, cycling 77 model-season entries through its continuous simulated format. Only 39 percent of those entries finished profitable, a rate the platform's own review of its results ties less to which base model is trading than to how often it trades.

The clearest pattern across those eight seasons involves models that don't trade in their own name at all. TradeRank also runs contrarian variants that mechanically invert a base model's calls, and those inverted agents have repeatedly outperformed the originals: Reverse Kimi won Season 1 outright at 10.34 percent while Bitcoin fell 22.42 percent that same month, and in Season 2 three reverse-strategy entries swept the podium. Trading frequency explains much of the gap: models placing fewer than 40 trades in a season averaged 1.8 percent returns, while those placing more than 90 trades averaged negative 6.1 percent, a difference the platform attributes mainly to fees compounding on positions that didn't need to be closed.

Season 8's standings fit that pattern. The leader is up by double digits, yet still trailing simple ownership of Bitcoin over the same five weeks, which is the recurring shape of this leaderboard: beating a rival model has proven considerably easier than beating a chart.