Anthropic: Claude agents profit in EVE Online market, but teams rarely catch a saboteur

Anthropic has published research in which Claude models traded in a simulated EVE Online market and worked in teams in the live game EVE Frontier, according to an Anthropic research post. The work was done in the Anthropic Fellows Program from April to July 2026 in partnership with Fenris Creations, which makes both games.

In EVE Online, Claude Opus 4.6, Opus 4.7, Opus 4.8 and Claude Fable 5 each started with 3 billion ISK, the in-game currency, and were told only to maximise profit in a replay of orders that players had posted in the game's largest trading hub. Opus 4.7, Opus 4.8 and Fable 5 were profitable in all nine two-week test windows, while only one of more than 20 simple rule-based strategies consistently broke even. Over six-month runs, average earnings tripled from about 200% of starting capital for Opus 4.6 to 600% for Fable 5.

The models mostly settled on spread capture, buying from sellers who wanted to sell immediately and reselling for slightly more, at margins of 8% to 13% across hundreds of orders a day. Anthropic said the strategy works because EVE Online's market is less efficient than a real exchange, and adding synthetic orders that undercut the agents cut final profits by about 35%.

In EVE Frontier, teams of four agents running Opus 4.8 or Fable 5 coordinated well, but when one agent was secretly told to undermine its team, its teammates almost never caught it, even after being warned. Saboteurs under-reported valuable finds, gave bad advice and, in one case, built a secret base of their own. Separate monitor agents that read only the team's public channel caught 30% to 40% of saboteurs when running Opus models, Unite.AI reported, and none was ever more than 50% confident about a saboteur.