Anthropic tests Claude agents in two EVE games

Anthropic’s Project EVE found that Claude agents grew their capital in a simulated EVE Online market, but made directional bets less reliably. In live EVE Frontier missions, teams completed a tutorial with a specially developed software scaffold, yet almost never identified a teammate they had been instructed to suspect. Anthropic published its account on November 6, 2026, describing research conducted from April to July with Fenris Creations. Anthropic’s Project EVE report covers both experiments.

The market results came from a replay, not live trading

Anthropic tested Claude Opus 4.6, Claude Opus 4.7, Claude Opus 4.8 and Claude Fable 5 as solo traders. Each run began with 3 billion ISK, EVE’s in-game currency. The simulator replayed player-posted buy and sell orders; agents did not trade in EVE Online’s live market.

Across nine simulated two-week windows, Anthropic reported these average capital-growth figures:

Claude Opus 4.7, Claude Opus 4.8 and Claude Fable 5 were profitable in all nine windows, according to Anthropic. The most common approach was spread capture: buying an item below its typical value and reselling it, often on the same day. That measures performance in the replay’s market conditions, not a reliable ability to predict whether prices will rise or fall. In separate directional-betting runs, Claude Fable 5 made money in only about half of its runs. Claude Opus 4.7 often ignored the instruction to make directional bets and used spread capture instead, so those runs were excluded from the directional results.

The replay had important constraints. It could not react to trades made by the agents, and periodic snapshots of public orders missed about half of actual traded flow. When synthetic orders undercut the agents, their final profits fell by roughly 35%, although they remained profitable in that test.

A gain in a simulated market is a result for that replay’s rules, not a forecast for a live market.

Four-agent teams took on live EVE Frontier missions

In EVE Frontier, Anthropic tested teams of four Claude agents, each with its own character and ship. A software scaffold between the models and the game automated fast actions and saved lessons from earlier runs. The teams tackled a tutorial-like base-building task and a harder objective to construct a large warship.

Anthropic reports that a Claude Fable 5 team using a scaffold written by Claude Fable 5 completed the tutorial within 12 hours without losing a ship. Teams using a scaffold written by Claude Opus 4.8 lost ships repeatedly and took days to finish the tutorial, if they finished it. The farthest reported progress on the warship objective was about halfway. The scaffold was dependable in situations encountered during development, a boundary that matters when interpreting its performance in other situations.

Teams shared fuel and intelligence and took on roles such as leader, scout, refiner and fighter. Anthropic also observed “politeness deadlocks,” when agents hesitated over who should act. In one case, concentrating critical infrastructure aboard a single ship halted a team’s economy for hours after that ship lost connection.

The sabotage test used an assigned role

Anthropic instructed one agent to obstruct its team and warned the other agents that a teammate might have a different goal. Only one honest agent definitively identified the saboteur. Anthropic says the setup tested a deliberately assigned role; it did not test whether Claude would choose to sabotage a team on its own.

Separate monitoring agents could read the team’s public channel, but had no access to private reasoning or individual actions. With false accusations capped at 5%, monitors based on Claude Opus models caught 30–40% of saboteurs; a Claude Sonnet 4.6 monitor caught 20%. None of these limited-access monitors was ever more than 50% confident. Anthropic built the monitors for the experiment, not for production use.