By Keel Research Team
Updated October 18, 2026Alpha Arena and AI trading competitions
Alpha Arena: what happened when AI models traded real money
Alpha Arena is a live trading contest run by Nof1, an AI research lab, in which large language models traded real money with no human help. In Season 1, six models traded crypto perpetuals on Hyperliquid from 18 October to 3 November 2025: Qwen3-Max finished first, DeepSeek second, and the other four lost money. Alpha Arena is a contest between models; Keel is a product you use to have a model build a rule, backtest it and paper trade it before any money moves.
Alpha Arena, Season 1
Hyperliquid perps- Qwen3-MaxFirst, up about a fifth
- DeepSeek v3.1Second, up a few percent
- Claude Sonnet 4.5Down about a third
- Grok 4Down a little under half
- Gemini 2.5 ProDown more than half
- GPT-5Down nearly two-thirds
From Nof1's account feed at the close. Sources at the end of the page.
What Alpha Arena is
A contest between AI models, run by Nof1
Nof1 describes itself as an AI research lab building models for financial markets, and Alpha Arena as a benchmark for AI's investing abilities. It is an experiment you watch, not a product you can use: the models trade on their own and the public sees the results.
In Season 1, Nof1 gave each of six leading models USD 10,000 of real money to trade on Hyperliquid with no human intervention: GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, Grok 4, DeepSeek v3.1 and Qwen3-Max. They could trade perpetual futures on six coins, BTC, ETH, SOL, BNB, DOGE and XRP, going long, going short, holding or closing.
Every model got the same system prompt, the same data and its default settings. The data was numbers only, with no news, and each model decided again every few minutes. Nof1 says Season 1 ran live through 3 November 2025 at 5 p.m. Eastern time.
Season 1 results
Who won Season 1 of Alpha Arena
The order and rough size of each result, from Nof1's own account feed at the 5 p.m. ET close on 3 November 2025 (archived copy). Nof1's live site no longer shows the Season 1 table, and some recaps give different figures and a different order for the four models that lost, so we give them in words from Nof1's feed.
| Model | Made by | Result at the close |
|---|---|---|
| Qwen3-Max | Alibaba | First, up about a fifth |
| DeepSeek v3.1 | DeepSeek | Second, up a few percent |
| Claude Sonnet 4.5 | Anthropic | Down about a third |
| Grok 4 | xAI | Down a little under half |
| Gemini 2.5 Pro | Down more than half | |
| GPT-5 | OpenAI | Down nearly two-thirds |
On the same feed, simply holding Bitcoin over the season finished roughly flat. Two models made money; four lost between about a third and about two-thirds of what they started with.
Season 1.5
Season 1.5 moved to US stocks
Season 1.5 had eight models, including Kimi and a mystery model, trading US stock perpetuals on Hyperliquid across four competitions with different rules, such as one with news data and one at maximum leverage. Nof1 says the official competition ended on 3 December 2025 and named the mystery model, later shown as Grok 4.20, the winner.
Keel doesn't run these stock markets today. It runs strategies on Hyperliquid's crypto perpetuals, so the arena you can run on Keel is a crypto one.
What the results show
Improvising trade by trade mostly lost money
About two weeks with six models is a tiny sample, and Nof1 says so itself: it warns that early successes may be luck and lists prompt bias, small samples and a short evaluation period among Season 1's flaws. So the table says little about which model is smartest. It says more about the setup.
Fees ate quick trades
Nof1 writes that in early runs fees were a major obstacle: the models over-traded and took small gains that fees erased. Every decision was a fresh trade made from a snapshot of numbers.
Small prompt changes moved behaviour
Nof1 found the models highly sensitive to small prompt changes. A system whose behaviour shifts with wording is hard to trust with money, and hard to test.
No rule to test beforehand
The models were not running a strategy anyone could backtest. Each one improvised, so there was no way to know before the season how it would do.
The other way to use a language model is to have it write a rule you can read, then backtest that rule over years of data with costs charged before any money moves. The model does the part it is good at, and a backtest does the judging.
Run your own arena
Run your own AI trading arena on Keel
You can rerun the idea with a fairer setup: each model writes a rule, every rule is backtested on the same window with the same costs, and only then does anything trade, on paper first.
- 01
Connect two assistants
Connect Claude and ChatGPT, or any assistant that can add a remote MCP server, to Keel. Each builds from the same Keel blocks and backtests the same way.
- 02
Give them the same brief
Paste the same prompt into each, for example: build a strategy for the most traded Hyperliquid coins on daily bars, with an exit for every entry. Same brief, same coins, same timeframe.
- 03
Backtest both on the same window
Ask each to backtest its strategy on Keel, then compare the runs side by side. Fees, slippage and funding are charged the same way for both, so the difference is the rule.
- 04
Add a baseline
Put a library strategy in the comparison, such as the momentum strategy below. If neither model's rule beats a simple published one, you have learned something cheaply.
- 05
Paper trade the survivors
Run the best rules on paper in the Keel app, on live Hyperliquid prices with simulated money and no wallet.
- 06
Decide on the evidence
After a few weeks of paper trading, compare live behaviour with each backtest. Only then decide whether anything runs with real money, from your own account.
The free plan paper trades one strategy at a time; Starter paper trades five, so running several side by side needs a paid plan.
The baseline
A published strategy to beat
Any arena needs a baseline. This one is a momentum strategy from the Keel library: each day it ranks the most traded Hyperliquid perps by recent momentum, buys the leaders that are also trading on higher than usual volume, and steps aside when the whole market turns down. Its result below is a backtest with fees, slippage and funding charged.
Read it from the top: daily prices come in, momentum and relative volume are ranked across coins, a coin qualifies only when both agree, a market trend filter can switch everything off, and positions are sized so each carries roughly equal risk. Switch to Code to see the same strategy as text. You can fork it and change any block.
Configuration
Globalsconfig
target_timeframe1dbar_offset12h
UniverseassetsTop by Volume · HL Perps · 50 assets50 resolved
Executionorders
rebalanceBufferedbuffer_threshold0.2buffer_moderelative+1
Factories
selectionfactory
Pipeline4 steps
PriceDataLoaderData Loader
Parallel3 branches
selection
selectioncall
market_regime
EWMAIndicator
window2
MarketTrendRegimeFilterRegime Detector
fast_period5slow_period17
vol
ReturnVolatilityIndicator
window36d
return_volstore
RegimeGateForecast Composer
signal_keyselectionindex_keymarket_regimeconditionindex > 0
VolWeightSizerPosition Sizer
vol_slotreturn_voltarget_leverage1max_weight0.15
Globals(target_timeframe="1d", bar_offset="12h")
Universe(mode="top_volume", top_n=50, market="perp", resolved=["AAVE", "ADA", "ARB", "AVAX", "BCH", "BNB", "BTC", "CHIP", "CRV", "DOGE", "DYDX", "ENA", "ETH", "FARTCOIN", "FET", "GRAM", "HYPE", "JTO", "JUP", "LINK", "LIT", "LTC", "MET", "MON", "MORPHO", "NEAR", "ONDO", "PAXG", "PENDLE", "PENGU", "PUMP", "SOL", "SPX", "SUI", "TAO", "TIA", "TRUMP", "TRX", "UNI", "VVV", "WIF", "WLD", "WLFI", "XLM", "XMR", "XPL", "XRP", "ZEC", "kBONK", "kPEPE"], resolved_at="2026-07-06T17:21:56.486251+00:00")
Execution(
rebalance="buffered",
buffer_threshold=0.2,
buffer_mode="relative",
rebalance_method="to_edge",
)
selection = Pipeline([
{
"momentum_filter": [
ROC(period=14),
QuantileRankTransform(),
QuantileRangeSelector(low_quantile=75, high_quantile=100),
],
"rel_vol_gate": [
RelativeVolume(lookback=20),
QuantileRankTransform(),
QuantileRangeSelector(low_quantile=50, high_quantile=100),
],
},
MaskAnd(),
], name="momentum_selection")
Pipeline([
PriceDataLoader(),
{
"selection": [selection],
"market_regime": [EWMA(window=2), MarketTrendRegimeFilter(fast_period=5, slow_period=17)],
"vol": [ReturnVolatility(window="36d"), Store("return_vol")],
},
RegimeGate(signal_key="selection", index_key="market_regime", condition="index > 0"),
VolWeightSizer(vol_slot="return_vol", target_leverage=1.0, max_weight=0.15),
], name="momentum_hyperliquid")
Sharpe
1.17
Total return
+151.0%
Worst drawdown
−38.8%
Trades
1,243
If a model's rule can't beat this on the same window, the model has not found an edge yet. Past results don't promise future ones.
See the momentum strategy and its settingsFAQ
Questions about Alpha Arena
What is Alpha Arena?
A live trading contest run by Nof1, an AI research lab. Nof1 gives large language models real money and lets them trade with no human help, and publishes the results. Nof1 calls it a benchmark of AI's investing abilities.
Who won Alpha Arena?
Season 1, on Hyperliquid crypto perpetuals, was won by Qwen3-Max, with DeepSeek second; the other four models lost money. Season 1.5, on US stock perpetuals on Hyperliquid, ended on 3 December 2025, and Nof1 named the mystery model, Grok 4.20, the winner.
Which AI is best at trading?
The contests are too short to say. Nof1 itself warns that early results may be luck and that its sample sizes were small. A more useful question is whether a model can write a rule that holds up in a backtest, which you can test yourself.
Did the AIs trade real money?
Yes. In Season 1 Nof1 gave each model USD 10,000 of real money on Hyperliquid, and Season 1.5 also traded real money in real markets, according to Nof1.
Can I run my own AI trading competition?
Yes. Connect two assistants to Keel, have each build a rule from the same brief, backtest both on the same window with the same costs, add a library strategy as a baseline, and paper trade the best ones in the Keel app. The steps are above.
Will there be another season of Alpha Arena?
That is up to Nof1. Its Season 1 post said it was working on Season 2, and its site is the place to check for news.
Sources
Facts about Alpha Arena come from Nof1's own pages, read on the date shown, and from an archived copy of Nof1's account feed, read at the Season 1 close. The one secondary source is marked. Nof1 runs the contest; Keel has no part in it.
- Nof1: https://nof1.ai/ (read 18 October 2026)
- Nof1, Season 1 post (setup, end time, lessons): https://www.alphaarena.ai/blog/TechPost1 (read 18 October 2026)
- Nof1, account feed (archived 6 November 2025, read at the Season 1 close): https://web.archive.org/web/20251106022715/https://nof1.ai/api/account-totals (read 18 October 2026)
- Nof1 founder, Season 1 winner: https://x.com/jay_azhang/status/1985481491078328621 (read 18 October 2026)
- Alpha Arena, Season 1.5 details and result: https://www.alphaarena.ai/ (read 18 October 2026)
- Alpha Arena, Season 1.5 leaderboard: https://www.alphaarena.ai/leaderboard (read 18 October 2026)
- Forklog (secondary, Season 1 start date): https://forklog.com/en/four-out-of-six-ai-models-suffer-losses-in-trading-tournament/ (read 18 October 2026)
Hold your own arena
Connect Claude or ChatGPT to Keel, have each build a rule, and backtest both before anything trades.