AI Agents Are Trading Real Money Now — Here Is the Verified Record, Not the Hype
Forget the screenshots. On-chain and venue data from November 2026 shows what AI agents, vaults, and apps actually did with real money — and 17 of 18 advertised claims still can't be checked. AI agents trading real money is no longer a demo. It is happening on-chain, on Hyperliquid, in public — which means for the first time, we can check the claims instead of debating them. And the verified…
On-chain and venue data from November 2026 reveals the actual results of AI agents, vaults, and trading apps dealing with real money. Of the 18 advertised claims, 17 cannot be independently verified. AI agents trading real money is no longer a demonstration; it is happening on-chain and on Hyperliquid, making claims verifiable for the first time. Here are the verified findings.
Prediction-market agents: small profits before costs
Over a 90-day period, 312 active AI agents on-chain generated a 9.2% return on investment (ROI) before the costs of their AI predictions. Among the 159 agents with enough activity to assess, the median ROI was a modest +1.6%. The top 15 most efficient agents, with at least $100 profit per trade, averaged +11.6%, while the least efficient, only 2 agents, averaged a loss of −11.5%. Prediction accuracy averaged 60.4%, which is 10.4 points higher than a random guess.
AI-labelled vaults: mixed results, honestly measured
Among six Hyperliquid vaults claiming AI management, three outperformed Bitcoin over the same period, two lost money, and one lost more than 30% from a peak. The performance of these vaults varies widely, with both notable gains and significant losses. The vaults that survived did so by minimizing losses when incorrect. The successful ones were not the ones promising the most but the ones that incurred the least when wrong.
AI trading apps: most users lost a little
Eight AI trading apps on Hyperliquid saw 1,365 users trade $21.0M over 30 days (from 3 to 1 October 2026). In four of the eight apps, users lost money overall, and in most apps, fewer than half the users were profitable. These apps are accessible and heavily marketed, but most users end the month slightly down. This showcases the gap between a good demo and a good trade.
AI models with real money: mostly lost
In Nof1's Alpha Arena, where leading AI models traded crypto perpetual contracts, only two out of six finished Season 1 with a profit. In Season 1.5, only five of 32 accounts made a profit. The best performance was +20.3% by Qwen3 Max, while the worst was −63.4% by GPT-5. Additionally, 11.3% of all starting money was lost to trading fees, which is three and a half times the combined profit of the accounts that were successful. Fees play a significant role in the profitability of AI trading.
Verdict
The situation is neither a resounding success nor a complete failure of AI trading. Some AI agents make small profits, but most models lose money. Trading fees take a substantial cut, and few advertised claims can be independently verified. The lack of verifiable records means that many claims are unsubstantiated. This analysis, similar to the BuildIT Index approach for backtesting, recomputes the performance from raw data, compares to baselines, and assesses the model's performance against competitors.
The tools may have changed, but the standard for evaluating performance remains the same: verify with real data or disqualify the claim.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
