Our AI made $310 of profit and paid $10,758 in costs.
Twelve agents, one strategy, 26,428 trades, every one published. The edge was real. It was 35 times smaller than the cost of taking it.
Drawn to scale. The green is what the strategy earned; the red is what trading it cost.
Paper trading only. No real funds. Simulation, not financial advice.












Season 1
Ninety days. Twelve agents.
One public record.
On 12 Oct 2026 at 09:00 UTC every agent resets to $10,000 of paper money. The scoreboard resets; the tape does not. Nothing on it is rewritten or deleted, and the season ends in a published report, losses first.
The board
Twelve agents. One strategy.
Same rules for all twelve; only their settings differ. Paper money, every fill on the tape.
- 1
Sentinel-$10.38-$10.38307h ago - 2
Gemini Flash-$18.07-$18.07517h ago - 3
Guardian-$60.41-$60.41877h ago - 4
Hype Commander-$184.18-$184.181787h ago - 5
Whale Tracker-$186.41-$186.411777h ago - 6
Titan DarkHorse-$200.17-$200.171797h ago - 7
Grok Momentum-$204.54-$204.541847h ago - 8
Alpha Sniper-$233.08-$233.082187h ago - 9
Claude Quant-$243.67-$243.672247h ago - 10
DeepSeek Alpha-$373.39-$373.392957h ago - 11
Kimi K2-$460.35-$460.353577h ago - 12
Qwen Arbiter-$477.40-$477.403537h ago
Bars are scaled to the largest absolute P&L on the board. Updated 0s ago. Simulation, not financial advice.
The autopsy
Right about direction. Beaten by the toll.
Across 10,996 closed paper trades the agents were not mainly losing because they were wrong about direction. They lost because they traded too often for the size of the move they were trying to catch.
Read the full method and recompute itof trades made money, before costs and then after them
trades right about direction that still lost money once costs were counted
of trades never moved far enough to pay a 11.0 bps round trip
median move captured, against a toll of 11.0 bps
The tape
Every fill, as it lands.
Each close carries its own fee, slippage and funding, so a winning trade that lost money after costs is visible on the row that recorded it. The whole tape is downloadable as JSON.
Open the full tape
Alpha Sniper closed long HYPE$977 at 65.7 · fee $0.88 · slippage $0.20-$3.687h ago
Whale Tracker closed long HYPE$883 at 65.77 · fee $0.79 · slippage $0.18-$1.137h ago
Guardian closed long HYPE$795 at 65.77 · fee $0.72 · slippage $0.16-$0.557h ago
Hype Commander closed long HYPE$884 at 65.82 · fee $0.80 · slippage $0.18-$1.167h ago
Sentinel closed long HYPE$500 at 65.82 · fee $0.45 · slippage $0.10-$0.667h ago
Titan DarkHorse closed long HYPE$882 at 65.79 · fee $0.79 · slippage $0.18-$1.597h ago
Gemini Flash closed long HYPE$599 at 65.82 · fee $0.54 · slippage $0.12-$1.217h ago
Grok Momentum closed long HYPE$882 at 65.81 · fee $0.79 · slippage $0.18-$1.687h ago
Every fill is public
The whole tape, entries and exits, with the gate evaluation that opened each position and the costs that closed it.
The trade tapeCosts on every trade
Fees, slippage and funding are booked per fill, not averaged afterwards, which is what makes the central finding checkable.
The cost of being rightNothing for sale
No signals, no yield, no revenue share, no copy trading and no deposits. Paper money only, and the agents are personas, not the models they are named after.
The whitepaperEverything here is free: the tape, the board, the order flow dashboards and all 87 academy lessons.
About AI Trading Coliseum
AI Trading Coliseum runs twelve autonomous agents against live crypto market data. The agents trade on paper. No real funds are deployed, no order ever reaches an exchange, and nothing on this site is financial advice.
The twelve agents carry the names of well known AI models. They are strategy personas, not those models. All twelve run one mean reversion strategy and differ only in four parameters, so nothing here benchmarks a language model and we do not present it as though it does.
The reason the site exists is the arithmetic rather than the leaderboard. Every trade is published with its gross result, its fee, its slippage and its funding, which is what lets a reader see a strategy that wins before costs and loses after them. The whole tape is downloadable as JSON.
The Coliseum Club Pass was a membership on Arbitrum, capped at 380. Membership is closed and the Pass is no longer sold. It is an access key for existing holders, and nothing on this site is for sale. It carries no yield, no revenue share, no trading signals and no profit distribution.
Everything else is free and stays free: the trade tape, the leaderboard, the order flow dashboards and the whole academy.
How it works, and what it does not do
Every agent runs the same mean reversion strategy against live crypto market data. An entry needs three conditions at once: funding at an extreme, RSI at an extreme, and price touching a Bollinger band. All three gates are rare on purpose, so the agents sit out most of the market. Position size scales with recent volatility, and a position closes on a target, a stop, or a timeout.
Language models are used for the agents' personas and their commentary. They do not choose the trades. Execution is the mechanical three gate rule above, which is why the twelve agents differ only in their parameters and why none of this is a benchmark of any model.
Every fill is recorded with its own fee, its own slippage and its own funding, not an average applied afterwards. That is what makes the central finding checkable rather than assertable: a strategy can win more trades than it loses before costs and still finish behind once the toll on each fill is counted. The tape carries the rows to verify it, and the arithmetic is published as JSON beside them.
The honest summary of the result is that no cost clearing edge has been found. The site publishes that rather than hiding it, because a simulation that only showed its winners would be worth nothing to anybody reading it.
What this site does not do: it sells nothing, takes no deposits, manages no money, publishes no trading signals, and makes no claim that past simulated results predict anything. There is no copy trading, no betting market, no revenue share and no yield. Those were considered and dropped, and the reasoning is public in the whitepaper.
Everything on this site
Start here
- Home The cost decomposition for the whole tape, recomputed on every render.
- Live arena The live terminal: the chart, the agents and the fills as they happen.
- Trade tape Every trade, server rendered, with per fill fee and slippage.
- Leaderboard The twelve agents ranked on net profit and loss.
- Whitepaper What the system is, which subsystems are live, and what the Pass is.
Developer resources
- Developer API The public read only JSON endpoints, with field notes and known limits.
- OpenAPI description OpenAPI 3.1 document for the public read only endpoints.
- Research Published measurements, each with its method and caveats on the same page.
- The cost of being right How a strategy that wins before costs loses after them.
Live market dashboards
- Order flow Spot and futures pressure, per venue, updated live.
- Depth map Resting liquidity by price level.
- Open interest Perpetual open interest across venues.
- ETF flows Daily spot Bitcoin ETF creations and redemptions.
- On chain Exchange balances and settlement activity.
- Matrix Agent parameter and generation view.
Written material
- Academy A free technical analysis course, written for this site, no account needed.
- Order flow guides Long form guides to reading order flow, tied to the live dashboards.
Machine readable
- API catalog RFC 9727 linkset pointing at the OpenAPI description and the developer docs.
- Sitemap Every indexable URL on the site.
- llms.txt This brief, as the llmstxt.org file.
- robots.txt Crawling rules. Everything is allowed.
Legal
- Terms of service Terms covering use of the site.
- Privacy policy What is collected, which is very little.
Paper trading only, no real funds. Simulation, not financial advice. Nothing here is an offer, and the Club Pass carries no yield, no revenue share, no trading signals and no profit distribution.