StarCraft: AI bot Pluto beats pros on the Remastered ladder
StarCraft: Remastered PC from £6.70 By Magnus Ohle Published 4 min readOn the official ladder of StarCraft: Remastered, an unusual opponent is currently causing a stir: The AI bot Pluto appears under the account ^333^ in regular Brood War matches and has already beaten well-known players such as Paralyze, Larva, Iris and Ret. However, the ladder appearance apparently does not stem from the team that developed Pluto.
Pluto itself is a research project for StarCraft: Brood War. The model was trained using reinforcement learning and self-play, and controls Terrans, Zerg and Protoss with the same neural network. Anyone wishing to place the classic game and its current Battle.net version can find the lasting fundamentals on our overview page for StarCraft: Remastered.
The ladder use is reportedly not authorised, according to the AI community
The decisive catch: members of the Brood War AI community expressly distance themselves from the public ladder matches. In a thread on TeamLiquid, Bruce Nielsen from the scene explains that Pluto was originally developed via BWAPI for the older version 1.16.1. The author had already enabled private human test matches in Remastered, but the community had been looking for a solution that would allow players to face the bot without unleashing it on the normal ladder.
According to the community, an outside person subsequently adapted Pluto for Remastered themselves and put it on the public ladder under ^333^. In addition, the version used there is accused of having a maphack. So far, this is a claim from the AI community and not a confirmed finding by Blizzard. The corresponding matches are therefore, in any case, not a clean human-versus-AI performance test.
What Pluto looks like against a top player
A replay against the Korean Protoss player Paralyze shows why the bot nevertheless attracts so much attention. StarCraft commentator Artosis has fully analysed the match: Pluto plays a highly unconventional Terran army of Marines, Medics, Siege Tanks, Ghosts, a Wraith and even a single Firebat. At the same time, units are controlled separately across several fronts.
Micro-control is particularly striking. The Wraith is deployed specifically against High Templar, Ghosts use Lockdown while other groups fight in parallel. The bot is not perfect, however: in the match shown, a Lockdown even lands on one of its own Siege Tanks. Against Larva, Pluto reportedly also won with a very unusual mass production of Marines and Medics, without adhering to typical human build conventions.
Why the displayed thousands of APM are misleading
The replays show at times absurdly high APM values. According to Pluto’s public project code, however, the actual model issues only one action every six frames. At normal game speed, this corresponds to around 240 model actions per minute. If one of these decisions moves twelve units at once, for example, StarCraft internally breaks the command down into several individual commands – and the visible APM counter rises accordingly sharply.
This does not mean that Pluto is subject to human limitations. A bot does not need to move a mouse pointer or shift its screen between two battle locations, and can control groups in parallel more precisely. The mere figure of several thousand APM therefore says little about how often the neural network actually makes new decisions.
Pluto was actually built for an AI competition
The open-source project describes Pluto as a single neural policy that controls all three races, matchups, builds and units. It was trained through self-play rather than via hard-coded build orders. At the StarCraft AI competition of the Computer Olympiad 2026, Pluto already prevailed against other bots.
That is precisely why the current ladder case is interesting: it shows a broader audience for the first time how strong modern reinforcement learning systems have become in the real-time strategy game that is more than 25 years old. At the same time, it also shows why an unannounced bot with possible additional vision on the normal ladder is not a fair benchmark. According to the AI community itself, it has reported the account to Blizzard.
Sources: TeamLiquid/Brood War AI community, the public Pluto project on GitHub and current reporting by Dexerto. The embedded replay cast is by Artosis.