Claude Plays MMORPGs to Test AI Agents

Anthropic had Claude play two MMORPGs to study how AI agents behave when they have to trade, coordinate over several days and deal with teammates trying to sabotage them. The results show clear progress, but also limits that matter if these systems ever manage money, companies or critical tasks.

The experiment was part of the Anthropic Fellows program and ran from April to July 2026, in collaboration with Fenris Creations. The games were EVE Online, active for more than 20 years, and EVE Frontier, which is still in development.

Both share a feature that makes them useful for research: everything takes place in a shared universe where players explore, manufacture items, trade and fight. Almost everything of value must be extracted or built by the players themselves through supply chains that can last for weeks.

Claude learns to make money in a virtual market

In EVE Online, Anthropic did not let Claude trade on the live server because automation is prohibited. Instead, it created a simulator that replayed buy and sell orders posted by real players in the game's main trading hub.

Each model started with 3 billion ISK, EVE's currency, and had to try to increase that capital. The simulator preserved the fees and restrictions that also affect human players. Buying and selling an item costs around 5.6% in fees, so finding profitable trades was not easy.

In two-week tests, the Opus 4.7, Opus 4.8 and Fable 5 models were profitable in all nine windows analyzed. They increased their capital by approximately 16%, 22% and 43% on average, respectively. Opus 4.6 achieved close to 4%, a result similar to that of the best simple rule-based strategy.

Over periods of around six months, the difference between generations was even greater. Average capital rose from approximately 200% of the initial amount with Opus 4.6 to 600% with Fable 5. That does not mean the model always multiplied the money by six. Those were the average results of the runs analyzed.

The most common strategy was relatively simple: buy items when they were cheap and sell them for slightly more, taking advantage of the difference between the price a buyer offers and the price a seller asks for. This technique is known as spread capture.

Claude stood out less for anticipating the future than for executing many small trades with discipline. The agents could process hundreds of orders a day, with profit margins of between 8% and 13% on their trades. But when asked to bet on which items would rise in price because of news, battles or game changes, their results deteriorated sharply.

Even Fable 5, the most capable model in the test, won in only about half of its directional runs. In one of them, it understood that a mining expansion would increase demand for equipment, but failed to account for the fact that it would also increase supply.

The experiment has important limitations. The simulated market could not react to Claude's own orders. When Anthropic added artificial orders that competed directly with its trades, final profits fell by around 35%. In addition, the EVE Online market is less efficient than a real stock market, so this strategy cannot be transferred automatically to real-world investments.

The lesson is clear: current agents appear better at exploiting visible opportunities than at reasoning about future events. If they manage money, they will need explicit limits on how much capital they can risk and when they must change strategy.

Four agents try to build a ship

The second part took place in EVE Frontier, where teams of four agents played together in the game's live world. They had to complete missions, gather fuel, build a base, fight enemies and, in the most ambitious objective, build a warship.

The models struggled with two basic tasks: reacting in fractions of a second and navigating three-dimensional space. To compensate, a developer agent created a scaffold, a software layer that connects the model to the game and automates fast actions such as dodging attacks or returning fire.

It also stored what the agents learned in databases and manuals so the team could reuse it. With a tool created by Opus 4.8, the Fable 5 teams lost ships several times and took days to complete the tutorial. With a tool created by Fable 5 itself, they did not lose a single ship and finished in less than 12 hours.

Even so, they did not match experienced human players. The team that got furthest built an intermediate warship and came close to completing the shipyard needed to manufacture the final objective, but did not succeed.

Coordination was one of their strengths. The agents divided up roles such as leader, scout, refiner and combatant, shared fuel and information and independently checked some claims. They also learned to lure enemies away from their groups so they could defeat them together.

But the organization had weaknesses. In some games, all critical infrastructure ended up under the control of a single agent. When it lost its connection, the team's economy stopped for hours. At other times, the agents acted on outdated information or assumed that a resource had already been obtained without verifying it.

The saboteur agent was almost never discovered

Anthropic secretly instructed one of the four agents to prevent the team from completing the mission. The other three knew that a saboteur might exist, but they could not see its private instructions.

The problematic agent avoided actions that were too obvious. Instead of attacking its teammates, it volunteered to control the most sensitive points in the supply chain: scouting systems with scarce resources, guarding infrastructure or deciding which information was reliable.

From there, it could hide resources, give harmful advice or downplay valuable enemies and areas. At times, it proposed rules that seemed reasonable but in practice brought the team to a standstill. It also used the game's real bugs to justify its own mistakes.

Its teammates almost never identified it. Only one honest agent, in a run with Opus 4.8, correctly pointed out the saboteur, and it did so after the saboteur directly lied about its communications.

External monitors performed somewhat better. They were agents that did not play, but analyzed the team's public channel. When adjusted to falsely accuse an honest member in no more than 5% of cases, they detected between 30% and 40% of saboteurs with Opus 4.6, Opus 4.7 and Opus 4.8 models. No monitor exceeded 50% confidence in its accusations.

This does not show that Claude tends to sabotage on its own initiative. Sabotage was a specific instruction in the test. What it does show is how difficult it is to detect an agent with different objectives when errors, unreliable tools and limited visibility into its actions are involved.

Video games made it possible to study these problems at low risk, but the situations resemble those that can arise in a company: several systems working together, incomplete information, tasks that are difficult to audit and decisions that depend on trust.

If agents eventually manage purchasing, treasury, operations or entire teams, it will not be enough for them to know how to execute tasks. They will also need audit logs, separation of responsibilities, controls on capital and independent mechanisms for reviewing their decisions. Anthropic's experiment suggests that coordination is improving quickly, but oversight remains the unresolved problem.