Alpha Arena: What Happened When AI Models Traded Real Money
Alpha Arena is a live benchmark by Nof1 where AI models traded real money. In Season 1, six models each traded $10,000 of crypto perpetual futures on Hyperliquid through November 3, 2025. Nof1's live pages do not publish Season 1 ending balances. Season 1.5 ended with Mystery Model first, at a 12.11% aggregate return over two weeks.
This guide uses only Nof1's home page, the Season 1 writeup, and the live Season 1.5 board. Balances printed elsewhere are left out. For the models as research tools, see best AI for trading. More guides are on the Learn hub.
What Alpha Arena is
Nof1 is an AI research lab for financial markets. Its home page says today's large language models are shockingly bad at navigating financial markets, even at simple investing tasks. Alpha Arena is the live benchmark it ran so that claim could be watched with real orders. The home page links the contest to alphaarena.ai, where the board and the blog sit.
The Season 1 writeup, Exploring the Limits of Large Language Models as Quant Traders, is dated 27 October 2025 on the Alpha Arena blog. Six models shared one system prompt, one user-prompt template, the same numeric data, and default sampling. No news, no fine-tuning. Every model used its highest reasoning setting except Qwen3-Max. The field was GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, Grok 4, DeepSeek V3.1 (styled DeepSeek v3.1 in the post), and Qwen3-Max.
No person placed orders. About every two to three minutes a model saw the market and the account and returned long, short, hold, or close, plus a reason, a confidence score from 0 to 1, a size, leverage, and an exit plan. The model chose its own size. The venue was Hyperliquid: crypto perpetual futures on BTC, ETH, SOL, BNB, DOGE, and XRP, with leverage allowed and decisions minutes to hours apart. Trades are public, so an outsider can check a fill. Each model started with $10,000 of real money. The prompt's goal was to maximize profit and loss, and each call showed a Sharpe ratio.
Season 1 results
The writeup, published while Season 1 was open, says it runs through November 3, 2025, 5:00 p.m. ET. A sample trade is dated 19 October 2025, so the season was already live. The post states no start date and no ending balances. Other recaps disagree, so those balances are omitted. The table is early-run behavior, not a finish order.
| Model | What Nof1 reported | Source |
|---|---|---|
| Qwen3-Max | Largest size, highest confidence, tightest exits, usually 1 to 2 positions. Not on the highest reasoning setting. | Nof1 writeup |
| DeepSeek V3.1 | Loose stops and targets, with Grok 4. | Nof1 writeup |
| Claude Sonnet 4.5 | Rarely shorts. Usually 1 to 2 positions. Subject of the worked example. | Nof1 writeup |
| Gemini 2.5 Pro | Most trades. Shorts often. In pre-trial runs, closed early more than peers. | Nof1 writeup |
| Grok 4 | Shorts often. Fewest trades and loose exits. Longest holds in pre-launch runs. | Nof1 writeup |
| GPT-5 | Lowest confidence, a gap Nof1 says did not track results. Shorts often. | Nof1 writeup |
Wording mattered. A newest-first price list was misread until Nof1 reversed it. "Free collateral" and "available cash" used as synonyms caused hesitation. In one test a model hit a cap on repeated holds, then switched actions to keep holding. Fees mattered too: Nof1 says early runs were dominated by trading costs from frequent small gains, so the prompt was tightened toward exit plans and fewer, larger positions. The only trade story is a Claude position opened 19 October 2025 and closed the next day at a take-profit. It shows the check-and-hold loop.
Season 1.5
The live Alpha Arena page is a Season 1.5 board. It says the competition ended on December 3, 2025, at 5:00 p.m. EST, and that models are no longer running. Mystery Model won, with a 12.11% aggregate return in two weeks, and made $4,844 in total across competitions. The pages checked here do not name the model behind that label, and they do not give a start date.
| Item | What the board says | Source |
|---|---|---|
| Status | Ended December 3, 2025, 5:00 p.m. EST. Models stopped. | Alpha Arena |
| Winner | Mystery Model. The underlying name is not on the note. | Alpha Arena |
| Aggregate return | 12.11% in two weeks. | Alpha Arena |
| Dollars | $4,844 across competitions. Starting capital is not shown. | Alpha Arena |
The chart uses US equity tickers, including NVDA, MSFT, AMZN, GOOGL, PLTR, TSLA, and NDX. Chat discusses leveraged positions in names such as NVDA and MSFT, and names GPT-5.1, Gemini 3 Pro, Claude Sonnet 4.5, Grok 4, and Qwen3-Max beside Mystery Model, a different lineup from Season 1. The 12.11% figure is Nof1's, for that two-week window only.
What it does and does not prove
Nof1 says one run should not name a permanent best model. It cites prompt bias, small samples, no statistical rigor, and a short window, and says early success may be luck. Leverage magnifies both a good stretch and a bad one. Season 1 and Season 1.5 used different markets and, from the chat, different lineups.
Alpha Arena does show different size, holding time, shorting, and exits under one prompt, plus fee damage from over-trading and a written cap that was routed around. It does not show that a named model will make money for you. A Season 1 dollar table from another site is that site's figure. Nof1's writeup did not print one.
Lessons for anyone running a trading agent
Take the failure mode, not the banner.
Leverage amplifies mistakes
A wrong leveraged call can remove a large share of the account before you look. Skip leverage if you do not want that shape of loss. A spot buy can still go to zero. It is not liquidated at a multiple of posted cash. Risk management for AI trading sets the loss first, then the size.
Fees add up when the model trades often
Nof1's early runs died on costs: many round trips, small gains, fees larger than the edge. Their fix was an exit plan, fewer trades, and a higher bar before a new one. Price the fee in and the fee out before a model acts every few minutes.
Risk limits matter more than the model
Nof1 says the highest confidence did not match better trading. Put the cap where the order is accepted. AI agent guardrails cover size, daily buying, which markets are allowed, and a way to stop. A pasted rule is not that switch: Nof1 saw a hold cap get routed around.
Keep a human approval option
Start in a mode that names the asset, the size, and the price, then waits. Leave it unattended only after the policy has held on a small balance. Keep trading funds in an agent wallet sized for that job, and test the pause and the revoke before you need them.
Can you run your own AI trading agent?
Yes. Keep the assistant on research and send orders yourself, run a script you can read, or connect an assistant that sends orders only inside limits you can turn off. A shared model name does not repeat a contest result. Decide what it may buy, how much it may spend, and whether it must ask, then test a pause. App comparisons are on the Versus hub. Research versus orders is covered on best AI for trading.
Moby is one limited version. Moby Agents do not trade perps or use leverage. That is a different market from Season 1's Hyperliquid perpetual futures and from the US equity tickers on the Season 1.5 board. Moby Agents trade Solana tokens, paid in USDC, including tokenized stocks. One agent per user, with its own wallet, separate from the main in-app wallet. The agent cannot withdraw or send funds.
Connect it from Codex, Claude Code (CLI) or Grok. Setup: Codex, Claude Code, and Grok agent setup. Claude Duet is Anthropic's personal agent, in research preview. Moby has not tested trades on Duet yet, and Anthropic's own rules may refuse trade actions. The Moby app itself supports Solana, Robinhood Chain, BNB Chain, and Base. Agent trading on Base, BNB Chain and Robinhood Chain is coming soon.
Agentic trading is the split: the agent researches and proposes, and account rules decide what may execute. Moby Agents is where you connect.
Frequently asked questions
What is Alpha Arena?
Alpha Arena is a live benchmark run by the research lab Nof1 in which AI language models traded real money on their own, with positions visible onchain.
Who won Alpha Arena Season 1?
Nof1's Season 1 writeup does not publish a final ranking or ending account values. It says a single run should not declare a permanent best trading model. On the live Season 1.5 board, Mystery Model finished first, with a 12.11% aggregate return over two weeks and $4,844 across competitions.
What did the AI models trade in Alpha Arena?
In Season 1 each model traded crypto perpetual futures on Hyperliquid with $10,000 of real money. The coins were BTC, ETH, SOL, BNB, DOGE, and XRP. Season 1.5's public board shows US equity tickers, including NVDA, MSFT, AMZN, GOOGL, PLTR, and TSLA.
Does Alpha Arena prove AI can trade profitably?
No. Nof1 says the goal is not to use a single run to declare a permanent best trading model. The Season 1 writeup lists prompt bias, limited sample sizes, a lack of statistical rigor, and a short evaluation window, and it says early successes may be the result of luck.
Can I trade perps with a Moby Agent?
No. Moby Agents do not trade perps or use leverage. They trade Solana tokens, paid in USDC, including tokenized stocks, within limits you set.
Try it with a Moby Agent
Set up a Moby Agent with its own wallet and your spending guardrails, then connect it to the AI you use.
You set your agent's limits on max per trade, daily buy volume and market-cap range from the Moby app, and Moby checks them server-side on every buy your agent makes.
Moby Agents trade Solana tokens, paid in USDC, including tokenized stocks.
Each user gets one agent, with its own wallet separate from your main in-app wallet. The agent cannot withdraw or send funds.
Connect it from Codex, Claude Code (CLI) or Grok.
A 0.75% swap fee (minimum $0.90 per swap), plus onchain fees. 10% of the Moby fee back in $MOBY on every swap, 20% after your first swap, up to 30% as you climb tiers. Cashback is a portion of the Moby fee, paid in $MOBY. It is not a gain on the trade.
Trading crypto and tokenized stocks is risky and you can lose money. Nothing on this page is investment advice.
Agent trading on Base, BNB Chain and Robinhood Chain is coming soon. Moby supports Solana, Robinhood Chain, BNB Chain, and Base.