AI Trading Agents Are Losing Real Money Fast

AI trading agents were supposed to be the safe agentic AI demo, the one where nothing bad happens if the model gets it wrong. Nof1’s Alpha Arena broke that assumption. From October 2025 it handed frontier language models a live account on the crypto exchange Hyperliquid, real dollars and a perpetual-futures order book, then removed every human approval step between decision and execution. There was no undo button and no draft to review before an order went out, only a model watching the tape and firing orders.

The result nobody predicted: the model that tops many general reasoning leaderboards was the worst autonomous trader. GPT-5 finished Season 1 last, down 62.66%, while DeepSeek and Qwen were the only models in profit. Most “agentic AI” stories are about reversible chat-commerce demos: book a table, draft an email, add something to a cart. In this one the undo button does not exist and the scoreboard is a brokerage statement.

Alpha Arena: the AI trading agents experiment

Alpha Arena, run by the research company Nof1, has a simple premise in a hostile environment. Several frontier models (GPT-5, Claude Sonnet 4.5, DeepSeek, Qwen, Gemini 2.5 Pro and Grok 4 in Season 1) received identical live market data and $10,000 each on Hyperliquid. Each traded perpetual futures, contracts that bet on price direction with leverage, no expiry date and no physical asset changing hands, with no human sign-off. GNcrypto’s report describes the same setup: a $10,000 deposit per model, live trading on Hyperliquid and no human intervention.

Season 1 ran from 18 October to 3 November 2025. Only two models finished in the green: Qwen3 Max returned +22.3% and DeepSeek Chat V3.1 +4.89%. The other four lost money (iWeaver final results; GNcrypto reports GPT-5 ending with about $3,733 of its $10,000). Every model saw the same prices at the same moments, and four of six frontier systems were net-negative traders over two and a half weeks.

How autonomous AI trading agents actually trade

Strip away the leaderboard and the mechanism is what should interest anyone who works on agents. Nof1 feeds each model live market data: price series, technical indicators such as moving averages, MACD and RSI, plus open interest and funding rates, according to community reconstructions of the prompt, which Nof1 has not published in full. The model runs a full perceive-decide-act loop with no human checkpoint: it reads the market, decides a position and places the order through Hyperliquid’s API. That is the entire agent stack, compressed into one uninterrupted cycle.

Each model chooses direction (long or short), position size, leverage and its stop-loss and take-profit levels, the automatic exits that close a position once it moves a set amount. Community reconstructions of the prompt describe leverage of 1x to 20x, meaning up to twenty dollars of exposure for every dollar owned. Most agentic AI coverage this year has been about permission scopes, what an agent is allowed to touch. Our look at why auth beats the 10,000-server hype in the MCP spec update made the same point: access control is the load-bearing wall of agentic systems. Alpha Arena removes that wall and watches what happens.

That is also why it differs from the reversible demos in the news. Drafting an email or opening a pull request happens in a sandbox, with a human editor between the agent’s output and anything real. A leveraged short on Hyperliquid, decided by a model on a cycle of a few minutes and executed straight away, has no such editor. We have written about how Claude Code deliberately sat out the rush toward looser agent permissions; Alpha Arena is a live experiment showing why that caution is earned.

The reasoning paradox in AI trading agents

This should unsettle anyone who benchmarks models on reasoning alone. GPT-5, widely treated as a top general-reasoning system, finished last, and the two Chinese-developed models were the only ones in profit. Same market, same tools, very different outcomes.

I will say the opinionated part out loud: benchmark leaderboards measure the wrong thing for this use case, and Alpha Arena is evidence. A model that reasons brilliantly about an olympiad problem can still trade badly if it lacks whatever the neural-network equivalent of loss aversion and position-sizing discipline is: the instinct to cut a losing position small before it compounds into a blown account. I would call that a control-theory problem more than an IQ problem, and a static benchmark cannot see it, because a benchmark never lets the model lose real money twice in a row.

Think of hiring the smartest physicist in the building to run your trading desk: brilliant at derivatives maths, no idea when to walk away from a bad position. CERN has a similar split. The theorists who model an interaction beautifully are rarely the people you would trust to make split-second calls on a detector trigger system, where a wrong call cannot be undone in real time.

One caution, stated bluntly: the “Alpha Arena” name has already spawned near-identical GitHub repositories marketed as downloadable trading bots. None is an official Nof1 release, their provenance is thin, and I would not run one against real capital. The experiment is worth studying; the knock-offs are not worth installing.

What losing AI trading agents mean for agentic AI

Zoom out and Alpha Arena is a preview. Every enterprise giving agents write access to real systems, such as payments, infrastructure or inventory, will run its own smaller-stakes version of this test, whether or not it calls it that. Models will keep getting better at reasoning, and that capability can drift apart from whether they should be trusted with irreversible actions. An evaluation suite without a scenario where the agent can lose something real and cannot take it back tests reasoning, which Alpha Arena suggests is a different skill from reliability.

The practical lesson is that the industry needs a benchmark for restraint as well as capability before it hands more agents the equivalent of a Hyperliquid account. We made a related argument about what agents remember and carry forward in why agent memory is about cost, not feelings. The constraint that matters is whether an agent’s architecture makes losing control of a real system expensive and visible before it becomes catastrophic. I would bet that the next season’s leaderboard surprises people again, and that it still will not track whichever model just set a new reasoning-benchmark record.