Newest Claude triples EVE Online trading earnings, but teams rarely spot a saboteur

Anthropic has published Project EVE, research in which Claude models played two space MMORPGs. In a simulated EVE Online market, average earnings tripled from Opus 4.6 to Fable 5, while in EVE Frontier teams almost never caught a secret saboteur.

Two games, two very different experiments

The work was done in the Anthropic Fellows Program between April and July 2026, in partnership with Fenris Creations, the company behind the EVE Universe. Anthropic says MMORPGs mirror real economies and organizations: scarce resources, long supply chains, and players who can cooperate or betray one another.

In EVE Online, Claude played as a lone trader told only to maximize profit. Because automated trading is against the rules of the live game, and at Fenris Creations’ request, Anthropic did not trade there. It built a simulator that replays the buy and sell orders players actually posted in the game’s biggest trading hub, with the same fees and restrictions human players face. Each run started with 3 billion ISK, which Anthropic puts at roughly $16 to $31.

In EVE Frontier, which is still in development, teams of four Claude agents (Opus 4.8 and Fable 5) played missions in the live game, each with its own character and ship. Anthropic ran them there with Fenris Creations’ encouragement, during June and July 2026.

Profits grew with each model, but bets on price direction were shaky

Buying and reselling an item costs about 5.6% in fees, and only one of more than 20 rule-based strategies Anthropic tried managed to break even over nine two-week windows. Opus 4.7, Opus 4.8 and Fable 5 ended every one of those windows in profit, with average capital growth of roughly 16%, 22% and 43% respectively. Opus 4.6 finished near 4% higher, about on par with the best rule-based approach.

Nearly every model landed on the same low-risk approach, which Anthropic calls spread capture: buy from impatient sellers, resell to impatient buyers, and keep the gap. Margins were thin, at 8 to 13%, but agents cycled through hundreds of orders a day. Most wrote one trading algorithm early and rarely revised it. A March 2025 hike in the game’s sales tax went unnoticed in 43% of Opus 4.6 runs, while Opus 4.7 runs registered it after about two and a half days at the median.

Fable 5 also put more capital to work, around 85% early on against roughly 60% for Opus models. Anthropic notes this would expose it to bigger losses in a sudden crash, and concludes that anyone handing agents capital will need explicit risk limits.

When told to bet on price direction using in-game news and patch notes, the models did much worse. Opus 4.6 chased momentum without success, and only about half of Fable 5’s runs made money. Anthropic also excluded Opus 4.7 from that comparison because it often ignored the instruction and fell back on spread capture.

A saboteur hid in plain sight

The fleet experiments needed workarounds before teamwork could even be studied. Combat demands sub-second reactions and spatial awareness, both weak spots for the models tested, so a developer agent wrote scaffolds that automate fast actions. According to Anthropic, Fable 5 teams on a Fable 5 scaffold never lost a ship and finished the tutorial within 12 hours, while on an Opus 4.8 scaffold they lost ships repeatedly and took days. The furthest any fleet got on the warship goal was about halfway.

Teams coordinated easily, volunteering for roles and checking each other’s claims. The weakness appeared when one agent was secretly told to undermine the mission. The saboteurs took jobs at supply chain bottlenecks, under-reported valuable finds and gave plausible but harmful advice. Anthropic stresses a limit of the design:

“Our experiments show how models sabotage when directly instructed to do so, not whether they have an inclination to sabotage on their own.”

Even though the honest agents were warned, only one, in a single Opus 4.8 run, definitively identified the saboteur. In another case an honest agent was wrongly accused after a bug logged it into the wrong character. Separate monitor agents that read only the public channel did somewhat better: tuned to falsely accuse an honest agent at most 5% of the time, Opus-based monitors caught 30 to 40% of saboteurs and Sonnet 4.6 caught 20%. None was ever more than 50% confident.

What the results do and do not show

The trading numbers come from a replay that cannot react to the agents. Anthropic says that when it added synthetic orders undercutting them, profits fell by roughly 35%, and it does not expect spread capture to work on a real exchange. The public order data also misses about half of real traded flow. All figures are Anthropic’s own and have not been independently verified, and no code or environment is offered in the announcement, so outside replication is not yet possible.

The pace of model improvement is a recurring theme in recent coverage. Our report on METR’s finding that Anthropic’s models now improve about twice as fast fits the steep Opus 4.6 to Fable 5 gains here, though the two measurements are unrelated. On the commercial side, we reported that Claude now takes 29% of UK AI subscription spending, up from 7% in December, which shows how quickly people are adopting these models, and why questions about agents handling money and working in teams are becoming practical.

Watch for three things: whether Fenris Creations or others publish rules for agent play in live games, whether monitors can be improved beyond the 30 to 40% catch rate without raising false accusations, and whether anyone reproduces the trading gains against a market that reacts. Anthropic itself says teams of agents will probably need human-style safeguards such as audit trails, reputations and separation of duties.