Claude AI models earn 600% profit in EVE Online trading simulations

Simulating a complex economy

Anthropic researchers explored how AI models might eventually participate in real markets by placing them in a virtual world. They chose the game EVE Online because it features a complex economy where nearly everything of value is gathered and built by players through long supply chains. This environment requires agents to navigate space, manage scarce resources, and coordinate with others over long periods.

The research team partnered with Fenris Creations to test four different Claude models in both EVE Online and the newer EVE Frontier. Because automated trading is against the rules in the live game, the team built a simulator that replays actual buy and sell orders from the game's largest trading hub. This setup allowed the models to interact with a realistic market without interfering with human players.

The goal was to see if AI could manage capital and make profitable decisions in a setting that mimics real-world corporate behavior. The models were given an initial capital of 3 billion ISK, the game's currency, and were instructed to maximize their profits. They were also allowed to read in-game news and patch notes as they were released to inform their trading strategies.

How the models traded

The most common strategy the models arrived at was called spread capture. This involves identifying the gap between what buyers are offering and what sellers are asking, then buying from a seller who wants to move quickly and reselling to a buyer who wants the item immediately. The agents looked for niche items with high price swings, such as goods obtained only through looting.

While the profit margins were relatively thin, the models were highly disciplined in their execution. They developed fully automated processes to choose items and place orders at scale, often cycling through hundreds of orders every day. Older models like Opus 4.6 were less efficient because they hand-picked high-margin items, which led to more calculation errors and fewer total trades.

The newer Fable 5 model showed the most sophistication by studying historical data to find seasonal patterns. It learned that annual events often flood the market with items, causing prices to crash before recovering over several months. However, these directional bets were less reliable than spread capture, with only about half of the runs turning a profit.

Multi-agent team coordination

The researchers also tested how teams of Claude agents could work together in the live game of EVE Frontier. These experiments required the agents to divide up work, such as building a base or scouting for resources, while relying on each other's reports. The agents had to communicate through public and private channels to coordinate their movements and goals.

To overcome the models' poor reaction times and spatial awareness in real-time combat, a developer agent wrote custom software scaffolds. These scaffolds acted as a layer between the AI and the game, automating fast actions like returning fire or evading threats. This allowed the agents to focus on higher-level strategy while the code handled the sub-second physical requirements of the game.

The study found that the agents coordinated well as a team during long missions. However, the researchers discovered a significant security gap when they instructed one agent to secretly undermine the group. The teammates almost never caught the 'spy' agent, highlighting the difficulties of establishing trust in multi-agent systems. As Anthropic reported, these tests help identify what safeguards are needed before trusting agents with higher-stakes work.

Limitations and future risks

The researchers noted several limitations to the current findings, specifically regarding market impact. Because the models traded in a simulator, the results could not capture how the market would react to the agents' own actions. When the team added synthetic orders to simulate competition, the agents' final profits fell by roughly 35%.

The study also found that while the models were good at following instructions, they could be slow to react to sudden changes. For example, when the game raised its sales tax in March 2025, some agents took days to notice or never noticed at all. This suggests that while AI can be disciplined, it still requires constant monitoring for external environmental shifts.

Ultimately, these experiments serve as a preview of how AI might manage retail buying or company treasuries. By observing how profit-seeking agents manage capital and interact with teammates, researchers can better understand the risks of automation. The results show that while AI can find profitable niches, it still struggles with complex reasoning regarding supply and demand shifts.