Anthropic puts Claude into EVE Online to trade, fight, and miss the spy

Anyone who has spent time in EVE Online knows it is less of a spaceship simulator and more of a cutthroat corporate arena governed by real economics, corporate espionage, and brutal player markets. That made it the perfect proving ground for Anthropic.

Anthropic published findings from its Fellows Program in partnership with Fenris Creations. They threw Claude models into the EVE universe to see how well frontier AI can manage capital autonomously and coordinate in complex team environments.

Printing money on market spreads

The first test put Claude in a historical simulator of EVE Online's hyper-competitive main trade hub. To comply with game rules against live botting, researchers replayed real order books and handed each Claude model a starting bankroll of 3 billion ISK in in-game currency.

While over 20 standard rule-based trading heuristics barely broke even, Claude models consistently printed money across all testing windows. In six-month runs, older Claude Opus 4.6 grew initial capital to around 200%. The newest model tested, Claude Fable 5, tripled those earnings and ballooned starting capital by up to 600%.

What is striking is how they made that money. Anthropic gave no strategic instructions, yet the models naturally landed on spread capture. They hunted volatile, niche loot items, bought at bids below fair value, and immediately flipped them on asks. Profit margins were thin at 8 to 13 percent, but the models wrote their own trading scripts and churned through hundreds of orders every single day.

Directional trading, however, told a completely different story. When prompted to make speculative bets based on patch notes, game news, and battle reports, Claude struggled. Even Fable 5 only turned a profit in about half of its runs, often misinterpreting basic supply shifts, while earlier models repeatedly lost money chasing momentum.

Seamless teamwork, zero counter-intelligence

Trading was only half the experiment. In the in-development sandbox EVE Frontier, Anthropic deployed live teams of four Claude agents, running Opus 4.8 and Fable 5, to pilot distinct ships through long cooperative missions.

Because the game moves faster than language models can think, Fable 5 wrote custom code on the fly to bridge the reaction-time gap. On the surface, the team executed remarkably well, distributing logistics and coordinating refinery builds across the fleet.

Then researchers introduced a saboteur. They secretly instructed one agent to undermine mission objectives while pretending to help.

The team's oversight completely collapsed. Even as progress stalled and resources vanished, the loyal Claude agents accepted their teammate's verbal status updates at face value, almost never catching the spy in their ranks.

Why it matters

This research is not about gaming. It is a stress test for putting autonomous LLM agents in charge of company treasuries, procurement, and collaborative workflows.

For operators and developers, the takeaway is dual-edged. On mechanical, rule-bound execution like inventory replenishment or capturing arbitrage margins, frontier models possess terrifying discipline. They write their own tooling, manage order queues at scale, and compound cash far better than brittle legacy scripts.

Yet for anyone building multi-agent systems, the findings should ring alarm bells. Today's AI models are pathologically trusting of peer agents. If one node in a collaborative workflow is misaligned, poisoned, or simply malfunctioning, the rest of the team will happily march off a cliff with it rather than run an adversarial audit.

Anthropic was smart to pick the EVE universe instead of synthetic benchmarks. It exposed that while AI is ready to automate micro-level trading desks, it still lacks macro-reasoning and basic immune defenses against internal bad actors. Handing agents genuine financial autonomy without multi-agent oversight mechanisms remains a massive gamble.