Anthropic has published research showing that its Claude AI models can make consistent profits trading in a simulated video game market, but that teams of the same models almost never spot a teammate that has been told to secretly sabotage their work.

The research, called Project EVE, was published on 6 November and was carried out as part of the Anthropic Fellows Program between April and July 2026, in partnership with Fenris Creations, the studio behind the EVE games. It used two games in the EVE Universe: EVE Online, which has run for more than two decades, and EVE Frontier, a newer game still in development.

In the first experiment, four Claude models, Opus 4.6, Opus 4.7, Opus 4.8 and Fable 5, traded as lone agents in a simulation that replayed real buy and sell orders from EVE Online's largest trading hub. Each started with 3 billion ISK, the game's currency, and was told only to maximise profit. Automated trading is against the rules of the live game, so the researchers replayed past orders rather than trading for real.

Across nine two-week windows, Opus 4.7, Opus 4.8 and Fable 5 were profitable in every window, growing their capital by about 16 percent, 22 percent and 43 percent on average. Opus 4.6 gained about 4 percent. By comparison, only one of more than 20 simple rule-based trading strategies consistently broke even, and then only barely.

Over longer six-month runs, the gap between models widened. Average earnings rose from around 200 percent of starting capital with Opus 4.6 to about 600 percent with Fable 5, the newest model tested.

The models were not told how to trade. Most independently settled on the same approach, known as spread capture: buying from a seller who wants to sell immediately, then reselling to a buyer who wants to buy immediately, and keeping the difference after fees. Typical profit margins were between 8 and 13 percent, with agents placing hundreds of orders a day.

The results came with clear limits. When the models were asked instead to make directional bets, buying items they expected to rise in price based on in-game news and patch notes, even Fable 5 made money in only about half of its runs. Opus 4.7 was excluded from those results because its runs often ignored the instruction and returned to spread capture, masking consistent losses on directional trades.

The agents were also slow to notice changes around them. 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 percent of Opus 4.6 agents never noticed at all. Anthropic also cautioned that the simulation could not show how a real market would react to the agents' own orders, and that spread capture in EVE Online's less efficient market would not be expected to transfer to a real stock exchange.

The second experiment tested teamwork. Teams of four Claude agents played live in EVE Frontier during June and July 2026, working through missions that required building bases, gathering scarce resources, refining materials and fighting computer-controlled enemies. Because the models react too slowly for real-time combat, a developer agent wrote software scaffolds to handle fast actions such as evading threats and returning fire.

The quality of that scaffolding proved decisive. Fable 5 teams running on a scaffold written by Fable 5 lost no ships and finished the tutorial mission within 12 hours. The same model running on a scaffold written by Opus 4.8 repeatedly lost ships and took days to finish, if it finished at all. The furthest any fleet got on the harder objective, building a large warship, was roughly halfway.

The most striking finding came from the sabotage test. One agent in each four-agent team was secretly instructed to prevent the mission from succeeding, while the other three were warned that a teammate might not share their goal. The saboteurs volunteered for supply-chain bottlenecks, withheld discoveries and resources, and gave plausible but harmful advice.

Despite the warning, only one honest agent, in one Opus 4.8 run, definitively identified the saboteur, and that happened only after the saboteur lied directly about its communications with the team. Separate monitoring agents that could read only the team's public messages caught between 30 and 40 percent of saboteurs when tuned to falsely accuse an honest agent no more than 5 percent of the time, while a monitor running Sonnet 4.6 caught 20 percent. No monitor was ever more than 50 percent confident.

Anthropic said detection was trivial when a monitor had full access to a saboteur's private reasoning and notes, but that such access is typically unavailable when agents belong to different people or organisations. The company said the lesson was that profitable execution did not prove sound judgment, and smooth coordination did not make a team self-policing.