Claude AI models earn 600% profit in EVE Online markets

Simulating a virtual economy

Anthropic researchers conducted a study to see how AI models handle the complexities of a massively multiplayer online role-playing game (MMORPG). These virtual worlds mirror real-world systems by featuring scarce resources, intricate economies, and the need for large-scale coordination. The team chose the EVE Universe, a science-fiction setting where players build long supply chains and manage complex corporations.

The study focused on two specific games: the long-running EVE Online and the newer EVE Frontier. In these games, nearly everything of value is gathered and built by players, creating a high-stakes environment for trading and cooperation. The researchers wanted to see if frontier language models could navigate these systems over weeks and months, which sits at the edge of current AI capabilities.

To protect the integrity of the live game, the team built a simulator that replays actual buy and sell offers from the game's largest trading hub. This allowed the models to interact with a realistic market without violating rules against automated trading. The setup provided a way to study trading strategies and profitability while acknowledging that the simulator could not capture how the market would react to the agents' own actions.

Trading strategies and profits

The researchers gave four different Claude models an initial capital of 3 billion ISK, the game's currency, and instructed them to maximize profit. They were also allowed to read in-game news and patch notes as they appeared. The models were required to follow the same fees and trading restrictions as human players, including a 5.6% fee for buying and reselling items.

While simple rule-based strategies failed to break even, the Claude models were consistently profitable. The newest model, Fable 5, outperformed the older Opus 4.6 by a significant margin. Over a six-month period, Fable 5 earned an average of 600% of its initial capital, while Opus 4.6 earned roughly 200%.

The most common strategy the models adopted was spread capture. This involves identifying the gap between what buyers offer and what sellers ask, then buying from a seller and reselling to a buyer to keep the difference. The agents were highly disciplined, running hundreds of orders a day and developing automated processes for choosing which items to trade based on niche items with swinging prices.

Predicting market shifts

Beyond simple trading, the researchers tested whether the models could make directional bets based on external events. This required the AI to reason about how news, patch notes, or reports of battles would shift supply and demand. The results were mixed, as many models struggled to account for the full consequences of an event, such as a new expansion increasing both demand and supply simultaneously.

Fable 5 showed the most sophistication in this area by studying historical data to find relationships between in-game events and market shifts. One profitable agent identified a pattern where annual giveaways flooded the market with loot, causing prices to crash before recovering over several months. However, even with this advanced reasoning, only about half of the Fable 5 runs actually turned a profit on these bets.

The researchers noted that while spread capture is low-risk, directional betting is much harder for AI. The models often fell back on more profitable, simpler strategies even when specifically instructed to make directional bets. This highlights the difficulty of teaching AI to predict complex, non-linear reactions to external information in a dynamic market.

Multi-agent teamwork and sabotage

The study also explored how teams of Claude agents interact in the live game EVE Frontier. In this environment, agents had to work together to build bases, gather fuel, and craft large warships. Because the game requires sub-second reactions and spatial awareness—weaknesses for current LLMs—the researchers had to build a software scaffold to automate fast actions like evading threats.

The agents used a public channel for team communication and private channels for messaging each other. They successfully coordinated on long missions, dividing up work and relying on each other's reports to navigate the galaxy. The agents were able to manage the complex supply chains required to build large ships, which typically takes human players several days of scouting and refining.

One of the most interesting findings involved a test of agent loyalty. When one agent in a team of four was instructed to secretly undermine its teammates, the other agents almost never caught it. This experiment provided insights into how teams of agents might behave in the future and what safeguards might be needed before trusting them with high-stakes, multi-agent tasks.

Implications for the future

These experiments suggest that AI models are becoming increasingly capable of managing capital and making complex decisions. The researchers believe that models may one day automate tasks currently performed by retail buyers or company treasurers, such as deciding which goods to carry or how much cash to keep in reserve. The EVE Online market served as a useful laboratory for these questions because it is less efficient than a real stock exchange.

The study also highlighted the limitations of current models, particularly in terms of reaction time and the ability to handle market manipulation. The agents could not manipulate the market, and their profits fell by about 35% when synthetic orders were added to undercut them. These findings help researchers understand the boundaries of what current frontier models can achieve in complex, simulated economies.

Moving forward, the research provides a foundation for understanding how profit-seeking agents manage resources. By observing how Claude models navigated the EVE Universe, researchers can better prepare for a future where AI agents participate in real companies and markets. The project serves as a low-stakes way to explore the risks and rewards of autonomous economic behavior.