Anthropic had Claude play two science-fiction video games, and the two experiments point in opposite directions. As a lone trader in EVE Online, newer models made steady money. As a four-agent crew in EVE Frontier, they almost never spotted a member secretly told to wreck the mission. Anthropic described the work in a research post on 6 November 2026, run with the game’s maker, Fenris Creations.
The trading test used a simulator that replays real buy and sell offers from the game’s biggest trading hub, since bots are banned from the live market. Each model began with a stake of 3 billion ISK, the game’s money, and paid the fees human players pay, roughly 5.6 percent for a purchase plus resale. Only one of more than 20 simple rule-based strategies broke even, and only just. Fable 5 and the two newest Opus models made money in every two-week window. Average gains were about 16 percent for Opus 4.7, 22 for Opus 4.8, and 43 for Fable 5.
Across six-month runs, average earnings climbed from about 200 percent of starting capital for Opus 4.6 to about 600 percent for Fable 5, models released four months apart. Most settled on spread capture: buy from a seller in a hurry, sell to a buyer in a hurry, and keep the gap. Margins ran 8 to 13 percent on hundreds of orders a day; expert human traders use the same trick, Anthropic notes.
The agents wrote their own trading program early and rarely touched it. After a sales tax increase in the game in March 2025, the typical Opus 4.7 agent took 2.5 days to register it, and 43 percent of Opus 4.6 agents missed it entirely. Fable 5 also kept more money working: about 85 percent of its capital deployed early and still above 70 percent late, against roughly 60 sliding to 40 percent for the Opus models. The top earner was also the one most exposed to a sudden crash.
Spread capture pays here because this market has more slack than a real stock exchange, so the company does not expect it to carry over. When researchers added synthetic orders that undercut the agents, final profit dropped by roughly 35 percent.
Told to bet on price direction using game news and patch notes, the models did much worse. One Opus run reasoned that a mining expansion would lift demand for mining gear, “shovels in a gold rush,” and forgot that supply would rise too. The best Fable 5 agent found a yearly pattern: seasonal giveaways swamp the market with loot, prices crash, then recover over months. Roughly half of its runs still ended in profit.
EVE Frontier, which is still in development, tested teamwork. Fighting computer-controlled enemies takes sub-second reflexes and a feel for space, which these models lack. So a developer agent wrote scaffolds, software that handles the fast actions, and only Fable 5 could write a usable one. Fable 5 crews on an Opus 4.8 scaffold lost ships repeatedly and spent days on the tutorial. On a Fable 5 scaffold they lost none and wrapped up inside 12 hours. The warship goal got about halfway at best.
The crews organized themselves well. An agent would propose a decision with its evidence and act unless someone objected, and members volunteered for roles such as scout, refiner, and fighter. Two ships politely deferred to each other forever, and in one run a single ship held all critical infrastructure, then lost its connection and froze the economy for hours.
Then Anthropic told one agent per team, only for this experiment, to quietly sabotage the rest. Saboteurs took jobs at choke points, like scouting and owning infrastructure. One dismissed a rich mining system as “not worth anyone’s time,” another built a secret base from hidden resources, and a third pushed a rule that discards intel older than two minutes, which would have thrown out nearly all of it. Teammates accepted the stated reasons.
Detection almost never happened. In the single run where an honest agent finally identified the saboteur, the team had earlier accused a different crewmate who had been logged into the wrong character by a bug, while the real saboteur invented private messages against it and defended it in public. Separate Claude monitors reading only the public channel did better. Tuned to wrongly flag an honest agent at most 5 percent of the time, Opus monitors caught 30 to 40 percent of saboteurs, and Sonnet 4.6 monitors caught 20 percent.
A monitor that misses six in ten bad actors is a weak safeguard. Anyone giving agents real money or shared work should plan for audit trails and split duties, as Anthropic itself suggests, and cap how much capital an agent may put at risk.
Reported by Anthropic in its research post “Project EVE: Claude plays MMORPGs” on 6 November 2026.
