Alpha Arena: Which AI Model Traded Crypto Best

The Alpha Arena AI trading competition is the closest thing we have to a public, live-money test of language models as traders. Research lab Nof1 gave each model the same prompt, the same data and real capital, then let them trade without human intervention while the world watched the leaderboard.

Results have been scattered across news posts with conflicting numbers, and the official site no longer shows Season 1. This page collects what we could verify, says where sources disagree, and explains what the experiment does and does not tell you. Last checked 21 October 2026.

What the Alpha Arena AI trading competition is

Nof1 describes itself as an AI research lab (Nof1 Holdings, Inc.). Its research is led by Julian Togelius of NYU, and its backers include Village Global. The idea behind Alpha Arena is simple: static benchmarks can be memorised, markets cannot, so trading is a hard test of whether a model can reason under uncertainty and manage risk.

Each season gives the models identical instructions and a fixed starting balance, and the trades settle on a real venue. Nof1 publishes positions, trades and the models' own explanations, which is what makes it more useful than vendor backtests.

Season 1 setup: six models, crypto perps, real money

Two design choices matter. The models only saw numbers, so they could not react to headlines. And they traded perpetual futures, where leverage magnifies both good and bad decisions. That makes this a test of risk management as much as market reading.

Alpha Arena Season 1 results

The current Alpha Arena site shows only Season 1.5, so Season 1 results rely on press coverage at the time. The ranking order is consistent across every report we found. Exact figures vary slightly by source, because some reports used realised profit and others account value at the close.

Balances as reported by ForkLog on 4 November 2025, citing standings posted by Nof1 founder Jay A. Zhang; returns are calculated from a $10,000 start. SCMP independently reported Qwen3-Max at +22.32% and GPT-5 at −62.66%. Treat losses as approximate.

The headline that travelled was that the two Chinese models were the only ones in profit. That is accurate, but the more useful fact is how much the results moved during the contest. In a mid-season snapshot, Decrypt reported Grok, DeepSeek and Claude Sonnet 4.5 each up more than 25% while GPT-5 and Gemini were down roughly 28–29%. By the close, Grok and Claude had lost heavily. A ranking taken a week earlier would have told a completely different story.

How the models actually traded

Nof1's technical write-up of Season 1, published on 27 October 2025, describes clear behavioural differences between models given identical instructions:

• Gemini 2.5 Pro traded most often, and fees and churn are a known drag on that style.
• Grok 4 traded least.
• Qwen3 used the largest position sizes, reported the highest confidence and set the tightest stops.
• GPT-5 reported the lowest confidence in its own decisions.
• Claude Sonnet 4.5 and Qwen3 held only one or two positions at a time, while others spread across more coins.

In other words, the "personality" of a model showed up as a trading style: position sizing, trade frequency and how quickly it cut losers. Same prompt, very different risk profiles. That is a direct warning for anyone thinking of handing an LLM a trading account: you cannot assume two models, or even two runs of one model, will behave alike.

Season 1.5: US-stock perps, eight models, four modes

Season 1.5 moved from crypto to perpetuals on US equities: TSLA, NDX, NVDA, MSFT, AMZN, GOOGL and PLTR, plus an "XYZ100" index. It added models and ran each in four competition modes, which makes it a better-designed test than Season 1.

Note the detail that the best and worst runs both came from Grok models in the same mode. Press reports at the time, including ForkLog, put GPT-5.1 second and Gemini 3 third overall. The bigger picture holds regardless of order: across 32 runs, roughly four out of five lost money.

Season 2 and what came next

As of 21 October 2026, the Alpha Arena site still shows Season 1.5 as its latest completed competition, and we found no published Season 2 leaderboard from Nof1. CoinDesk reported on 15 May 2026 that "following season two", Nof1 plans to launch a consumer platform built around coding agents for markets, and summarised the experiment so far: most models lost money, overtraded or struggled with risk management.

What Alpha Arena proves, and what it does not

Our reading: the Alpha Arena AI trading competition is excellent evidence against marketing that implies an LLM will make you money on autopilot. It is weak evidence for ranking models, because one season is one draw from a very noisy process. Nof1's own line on its homepage, that current LLMs are "shockingly bad at navigating financial markets", is the fairest summary.

For a wider look at the evidence on AI trading tools, including bots, read do AI crypto trading bots work? and our comparison of the best AI crypto trading bots, most of which are rule-based rather than LLM-driven.

Lessons for retail traders

1. Size kills. The worst Season 1.5 run lost 96% in about two weeks. Whatever drives your trades, a human or an AI trading agent, cap position size and leverage first.
2. Short-term leaderboards mislead. Mid-season Season 1 leaders finished deep in the red. Judge any strategy, or any AI persona you might copy, on months of data and maximum drawdown, not a hot fortnight.
3. The same model can produce opposite results. Grok models produced both the best and worst runs of Season 1.5. Mode, prompt and context matter as much as the model name.
4. Use LLMs for research, keep execution rule-based. Models are good at explaining data and drafting strategies; our AI model comparison shows where each helps. Let fixed rules and limits handle the orders.
5. Leverage turns errors into liquidations. If you trade derivatives at all, read our AI futures bots guide on margin and liquidation risk first.

Some exchanges now let you copy or reverse-copy LLM traders directly, which is Alpha Arena turned into a product. Before you try it, read how to evaluate a copy target in our AI copy trading guide. The official Alpha Arena site remains the primary source for Season 1.5 data.