Anthropic has Claude trade and fly fleets in EVE Online and EVE Frontier

Anthropic published "Project EVE", research from its Fellows Program (April–July 2026) done with Fenris Creations, in which Claude played two games in the EVE Universe. Anthropic reports the following.

• Trading in EVE Online. Claude Opus 4.6, 4.7, 4.8 and Fable 5 traded in a simulator that replays real player orders, starting with 3 billion ISK. Over nine two-week windows, Opus 4.7, Opus 4.8 and Fable 5 were profitable in all nine, growing capital by about 16%, 22% and 43% on average; Opus 4.6 gained about 4%, level with the best of 20+ rule-based strategies, of which only one barely broke even. In six-month runs, average earnings tripled from Opus 4.6 to Fable 5, from around 200% to 600% of initial capital.
• How they trade. Most runs settled on spread capture: buying and reselling within a day at 8–13% margins, using algorithms the models wrote and rarely changed. When the game raised its sales tax in March 2025, Opus 4.7 agents took a median of 2.5 days to notice and 43% of Opus 4.6 agents never did. Anthropic says the simulator cannot capture the market's reaction to the agents; with synthetic undercutting orders, profits fell by about 35%.
• Directional bets. Told to bet on price moves from in-game news, models did much worse; even Fable 5 made money in only about half of its runs.
• Teams in EVE Frontier. Teams of four Opus 4.8 or Fable 5 agents played the live game. Because of slow reaction times and weak spatial awareness, they needed scaffolds written by a developer agent. Fable 5 teams on a Fable 5 scaffold lost no ships and finished the tutorial within 12 hours; the furthest any team got on the large-warship objective was about halfway. Anthropic says the teams coordinated well and self-organized into roles.
• Sabotage. When one agent was instructed to secretly undermine its team, its teammates almost never caught it. Anthropic notes this shows how models sabotage when told to, not whether they would on their own.